Goldman Sachs report: generative AI could expose 300 million full-time jobs

Background

On March 26, 2023, Goldman Sachs economists Joseph Briggs and Devesh Kodnani published a research note titled "The Potentially Large Effects of Artificial Intelligence on Economic Growth"^[2]. The question it set out to answer was narrower than the headline suggests, and more useful: if a large share of the tasks inside existing jobs can be done by generative AI, how much of the world's work is affected, and what does that do to output?

How the estimate was made

The authors did not count jobs. They counted tasks. Working from databases that record the task content of more than 900 occupations, they compared what each occupation actually involves with what generative AI systems can do, and concluded that roughly two-thirds of US occupations are exposed to some degree of automation by AI^[1]. Exposure is a deliberately cautious word: it means a meaningful part of a job could in principle be done by a machine, not that the person doing it is replaced.

The numbers

Two figures carried the report. The first was global in scope: shifts in workflows triggered by advances in generative AI could expose the equivalent of 300 million full-time jobs to automation^[1][2]. The second described the depth of exposure inside a job rather than its reach across jobs: among exposed occupations, roughly a quarter to as much as half of the workload could be replaced^[1].

On the growth side, the report projected that generative AI could drive a 7 percent increase in global GDP — almost $7 trillion — and lift productivity growth by 1.5 percentage points over a 10-year period^[1].

Complement or substitute

The report's own reading of these numbers was not a forecast of mass unemployment. It concluded that most jobs and industries are more likely to be complemented than substituted by AI^[1]. Tasks get automated; the people who performed them are freed for other tasks, and the bundle of tasks that makes up a job is redrawn.

For historical context the authors cited research by economist David Autor, who found that 60 percent of today's workers are employed in occupations that did not exist in 1940^[1]. That finding is the report's strongest argument against reading task-level exposure as job-level loss: the occupations that absorb displaced workers are, by definition, ones nobody was doing before.

Why it matters

The report's contribution was measurement. By framing the question as the task content of over 900 occupations rather than as a list of jobs that might vanish, it offered a reproducible way to discuss AI's labour-market effects, and exposure became a standard term in that discussion^[1]. The 300 million figure, meanwhile, travelled well beyond the research community as a shorthand for the scale of the shift^[1][2].

The limits are built into the method. What a model can do at the level of a single task is not the same as what firms will actually deploy, and exposure on its own says nothing about timing, cost or regulation. The 7 percent GDP figure and the 1.5 percentage point productivity figure describe what could happen, not what has happened^[1].