AI Will Only Take 5% of Jobs, According to a Nobel Laureate?

The sentence that keeps coming back

"Solo il 5% di noi perderà il lavoro per colpa dell'AI." That is the headline of Punto Informatico, an Italian tech outlet, on October 7: only 5% of us will lose our jobs because of AI. The same day, India's Business Today ran "'AI won't take your job anytime soon': Why Nobel laureate sees only 5% of jobs changing in 10 years."

The trigger came a day earlier. Mustafa Suleyman, who runs AI at Microsoft, reposted on X an essay by Daron Acemoglu, the 2024 Nobel economics laureate. It appeared in July in The Humanist Review of AI, the magazine Microsoft AI publishes. And the line isn't new: on October 2, 2024, Bloomberg's headline was already "AI Can Only Do 5% of Jobs, Says MIT Economist Who Fears Crash."

The figure has a public source: Acemoglu's research paper "The Simple Macroeconomics of AI", circulated by the NBER in May 2024 and later published in Economic Policy. We read it. The number is in there, give or take a few tenths. The unit isn't.

4.6% of what, exactly

The calculation fits on a single line: 0.23 × 0.200 = 4.6%.

The 20% is the share of US labor tasks "exposed" to AI, taken from a 2023 study by Eloundou and coauthors, aggregated by occupation and weighted by wage bill. The 23% is the share of those exposed tasks that would be profitable to hand to a machine within ten years, an estimate borrowed from work on computer vision. The result: 4.6% of tasks, rounded to 5%.

Two details change the reading. First, the author says outright that he takes these exposure measures "without distinguishing automation vs. task complementarities". "Exposed" means AI can step into the task, not that it takes the place of whoever does it. Second, he warns that neither factor can be estimated accurately, and asks that his numbers be read "with great caution as suggestive numbers".

A job is more like a key ring. Knowing that a machine can open one lock in twenty doesn't tell you how many people will hand back their keys.

What the paper concedes, and what headlines do with it

The line ends with a parenthesis, though: "4.6% of all tasks (or occupations)". The paper itself nods toward jobs, because exposure is measured occupation by occupation. But it never turns a wage-weighted share of tasks into a count of eliminated positions. Nothing in the calculation counts layoffs.

At Bloomberg, the headline slips more than the text. In the 2024 piece as carried by its Portuguese edition, Bloomberg Línea, only 5% of jobs can be replaced, or at least heavily helped, by AI over the next decade. That nuance fits the exposure measure. The headline lost it.

Punto Informatico goes furthest: its body says 5% of human beings will be replaced by AI in their working lives within ten years. Business Today first quotes the essay correctly, 5% of what humans do, then writes a paragraph later that only 5% of human jobs can be replaced by AI. Even the MIT center that Acemoglu co-directs sums up the Bloomberg interview as being about "5% of jobs".

The 0.66%, a ceiling read as a forecast

Those 4.6% of tasks, multiplied by average cost savings of 14.4% on the affected tasks, give 0.0066. That is a total factor productivity gain of at most 0.66% over ten years, about 0.064% a year.

"At most" is in the text. The paper says the effect "should be no more than 0.66% in total", and its conclusion calls it "an upper bound".

It then lists what could make even that too high. Gains from digital technologies often follow a J-curve, flat for years before they climb. If so, Acemoglu writes, "the 14.4% overall cost reductions may be a significant overestimate for the next 10 years." And once tasks that are hard for a machine to learn are factored in, the bound drops to about 0.53%.

A bound isn't a wall, though. The same paper prices more generous variants: 0.89% with productivity gains taken from another study, and about 0.9% if automation becomes profitable on 30% of exposed tasks. The 0.66% is the ceiling of his baseline scenario, under assumptions he names one by one. Think of it as the speed limit on the sign, not the arrival time on your GPS.

On May 31, 2024, Le Monde Informatique, in French, had Acemoglu saying that AI would raise productivity by 0.66% over ten years, in quotation marks. The paper's sentence said the effect should not exceed that level. The difference is a few words, and it separates a ceiling from a forecast. The same article reports the lowered 0.53% bound correctly, and two days earlier The Register had quoted the exact wording: "should be no more than 0.66 percent in total".

When the author rounds it himself

The drift doesn't come only from the press. On August 6, 2024, in a signed Fortune op-ed, Acemoglu writes that you come up with a productivity boost of "only 0.66%" over ten years, and that the effect is "more likely to be in the range of about 0.5%-0.6%". He concedes "a huge degree of uncertainty", but the ceiling has become a likely range.

In the July 2026 essay, he tells the story differently again. The 5% becomes a quarter of the 20% of work that, he writes, "can be fully automated", where the paper spoke of exposed tasks, automation and assistance lumped together. He calls his 2024 numbers "my forecast" and the 5% "no more than a guesstimate".

The GDP figure slides the same way. The essay recalls a forecast of "about 1.5%" more GDP over ten years. The paper gave two ranges: 0.93% to 1.16% if investment follows modestly, 1.4% to 1.56% in a boom. So the 1.5% is the top of the most favorable scenario. And it looks like the Goldman Sachs number the paper takes on, except that one was about 1.5% more productivity growth every year.

What still holds

The thesis stands up. On his own assumptions, Acemoglu shows we are far from the economic revolutions others have promised. In 2026 he judges that his numbers are holding for now, while admitting that models have improved faster than he expected.

What the phrase "5% of jobs" adds, however, appears in none of these calculations: not how many positions disappear, not how many get transformed. We saw this in April with the Coface study: AI slices jobs into tasks long before it eliminates any. And the J-curve Acemoglu invokes echoes the productivity paradox we covered in March.

The paper counts tasks and sets a ceiling. The headlines count people, and announce a future.