Goldman Sachs's "300 Million Jobs" Estimate — What "Exposed" to Generative AI Meant

The "AI could replace 300 million jobs" headlines came from a report dated 26 March 2023 (US Eastern Time) by Goldman Sachs's economics research team, "The Potentially Large Effects of Artificial Intelligence on Economic Growth". The economists Joseph Briggs and Devesh Kodnani estimated that shifts in workflows triggered by generative AI could expose the equivalent of 300 million full-time jobs to automation.

The key word is "expose". The same report said most jobs and industries were only partially exposed and therefore more likely to be complemented than substituted, and under its illustrative assumptions only 7% of US employment would be substituted by AI. The 300 million was not a count of people out of work; it was the volume of work AI might take on, expressed as full-time jobs.

What was estimated, and how

The report appeared four months after ChatGPT launched and about two weeks after GPT-4 was announced. The method:

1. Examine the task content of more than 900 occupations in O*NET, the US Department of Labor's occupational database (later extended to more than 2,000 occupations in the European ESCO database).
2. Based on existing literature on the likely uses of generative AI, classify 13 of O*NET's 39 work activities as exposed to AI automation.
3. In the base case, assume AI can complete tasks up to a difficulty of 4 on O*NET's 7-point scale.
4. Assume occupations that spend much of their time outdoors or on physical labour cannot be automated by AI.
5. Take an importance- and complexity-weighted average of each occupation's tasks to estimate the share of its workload AI could replace.

The results:

As the last row shows, changing the assumed difficulty of tasks AI can perform, or allowing AI to work with robotics, moves the estimate anywhere from 15% to 35%. The report described its baseline as on the conservative side of existing estimates.

"Exposed" is not "lost"

The report illustrated three groups. Jobs where 50% or more of importance- and complexity-weighted tasks are exposed were assumed likely to be substituted; 10–49% likely to be complemented; 0–9% unlikely to be affected.

The report cautioned that these shares would depend on how labour demand and workloads respond to partial labour savings. It added that worker displacement from automation has historically been offset by new jobs. Goldman's summary article cited research by the economist David Autor finding that 60% of today's workers are employed in occupations that did not exist in 1940.

On productivity, the report estimated that generative AI could raise annual US labour productivity growth by just under 1.5 percentage points over the ten years after widespread adoption, and eventually raise annual global GDP by 7%, almost $7 trillion. It also noted that the timing of any such boom is hard to predict.

A parallel estimate

On 17 March 2023 a paper by researchers including OpenAI staff, "GPTs are GPTs" by Eloundou, Manning, Mishkin and Rock, was posted to arXiv. It found that around 80% of the US workforce could have at least 10% of their tasks affected by large language models, and about 19% at least half, and stated that it made no predictions about adoption timelines. Like Goldman's work, it measured the share of tasks that could be affected, not changes in employment.

Where Frey and Osborne in 2013 sorted whole occupations into automatable or not, both 2023 estimates measured the share of tasks within occupations. That made it possible to separate substitution from complementarity.

Criticism, and later numbers

The scale of the estimate was disputed. MIT economist Daron Acemoglu, in "The Simple Macroeconomics of AI", estimated a much smaller GDP gain of 1.1–1.6% over ten years. According to MIT News, he said AI would affect a set of office jobs about data summary, visual matching and pattern recognition that amount to about 5% of the economy, and called the likely productivity gain disappointing relative to the promises made by the industry and tech journalism.

Goldman Sachs updated its own numbers on 13 August 2025. If AI were widely adopted, it said, 6–7% of the US workforce could be displaced, but the impact would likely be transitory as new jobs emerged, with unemployment rising by about half a percentage point during the transition. If current AI use cases were expanded across the economy, 2.5% of US employment would be at risk. Briggs and Sarah Dong wrote: "we remain skeptical that AI will lead to large employment reductions over the next decade." The same note pointed out that unemployment among 20- to 30-year-olds in tech-exposed occupations had risen by almost 3 percentage points since the start of 2025. Research using payroll data on young workers is on the "canaries in the coal mine" page.

A view from Japan

In July 2024 Japan's Cabinet Office devoted the first chapter of its half-yearly world-economy report to how AI is changing labour markets, organising international research around the distinction between AI that replaces human work and AI that assists it. The substance of that review is covered on the Nomura Research Institute page, which deals with Japan's own 49% estimate from 2015. The argument over 300 million jobs was, in the end, an argument over where the line between substitution and complementarity falls.

The wider sequence of events is on the AI and society timeline.