New Delhi: Artificial Intelligence could make the US economy about a third larger by 2030 and still leave nearly one in five office and professional workers without a job, according to a study by economists at The Anthropic Institute.
That figure, according to the study, ‘Economic Scenarios for Transformative AI,’ comes from the most disruptive of three scenarios. It says that the scenarios “are not predictions”, with the authors attaching “no probabilities to them”.
Their stated aim is to show what the economy would look like under different assumptions about how far and how fast AI spreads.
Anton Korinek, an economics professor at the University of Virginia’s Department of Economics and Darden School of Business, who also sits on Anthropic’s Economic Advisory Council; Charles I. Jones, the STANCO 25 Professor of Economics at Stanford Graduate School of Business; and Szymon Sacher, Tess Cotter and Peter McCrory of Anthropic’s economics team are the authors of the paper.
The Anthropic’s research arm released the report Wednesday.
The model splits workers into two groups. One is what the study calls cognitive occupations, meaning management, professional, sales and office jobs, the desk work that AI can take on. The other covers everything else, such as construction and electrical work, which the study assumes AI leaves untouched.
It then asks how jobs, pay and growth change as AI does more of the desk work. Each number is measured against a “no-AI path,” an estimate of where the economy would have gone if AI had never arrived. For instance, a figure quoted as 8 percent above the no-AI path means 8 percent higher than that AI-free baseline.
The three scenarios
Under the first scenario, which the authors call ‘modest’, output in the year 2030 sits 1.6 percent above the no-AI baseline and unemployment moves from 3.8 to 3.9 percent. The paper says this case “has only small labor market effects”.
In the ‘substantial’ scenario, output runs 8.3 percent higher and the economy grows about 5.4 percent a year, against a normal rate of 2 percent. The paper notes for comparison that the “fastest GDP growth rate during the 1990s dot-com boom was 4.7 percent in 1999”. Pay for office workers roughly stalls, while pay for other workers rises about 6 percent.
As for the ‘extreme’ scenario, annual growth reaches 15.4 percent—fast enough to double average incomes every five years, “compared with doubling every 35 years in the US over the past century”.
The share of office and professional workers with jobs falls by about a fifth. Joblessness among those workers reaches 17.9 percent, which, the paper calls, “far above any postwar rate in the United States”. Unemployment across the whole economy rises to 11.9 percent, higher than the roughly 10 percent seen after the 2008 financial crisis.
Pay diverges in this scenario. The average wage rises, but the average conceals a split. Office and professional pay falls about 11 percent, while pay in other jobs rises about 34 percent, as those workers become harder to find and demand shifts toward them.
Also Read: GPT-6 Astra is here. What OpenAI’s new model can do
Bigger economy, smaller share for workers
The study tracks how income is divided between workers and the owners of business assets, meaning factories, equipment, computers and the data centres that run AI. At present, of every dollar the economy produces, about 60 cents reaches workers as wages and 40 cents goes to those owners. As AI takes over tasks, more of the dollar flows to the owners.
The shift is widest under the ‘extreme’ scenario. Workers’ share drops to 45 cents and the owners’ share climbs to 55 cents. Because the economy grows by a third while workers’ slice shrinks by a quarter, the study finds total wages in 2030 end up “almost exactly what it would have been without AI”. Nearly all the additional output goes to the owners of assets, whose income rises about 80 percent.
The authors then turn to compensation. Since the economy grows so much in the third scenario, the winners could in theory cover the losers. The total gain runs close to three times what office workers lose in pay and jobs.
Covering that loss in full would need a transfer of “about 9 percent of GDP, roughly the size of Social Security and Medicare combined”. Transfers on that scale, the authors add, “have no precedent”, and after past shocks such as competition from Chinese imports, this kind of support “mostly does not happen on its own”.
The paper also flags what it calls “so-so automation,” i.e., AI that pushes people out of work without making the economy much more productive. In that situation, wages can fall outright, and “compensating workers would consume most of the gains”.
What Americans expect
To understand public opinion, researchers surveyed 10,980 American adults in August. They asked when AI might be able to perform eight tasks of increasing difficulty, how widely it would be used, how much time it could save, whether it would work alone or with a person, and how long laid-off workers might take to find new jobs.
The median respondent’s answers suggest the substantial scenario, with output rising about 8.6 percent and unemployment reaching around 4.6 percent.
By 2030, 52 percent of respondents thought AI could run an online business with no staff, 39 percent expected a Nobel-level discovery, and 40 percent said such a discovery would never happen.
The authors also explain what the study does not cover—such as catastrophic risks, political factors, business cycles or possible financial market disruptions. It excludes robots and physical labour, which is why the analysis ends in 2030.
The study also assumes that workers who lose jobs can move to a new field and earn the going wage there, without accounting for the loss of skills from their previous jobs.
Reviewers were divided. Some said the extreme case “is better read as a thought experiment,” while others felt the modest case “understates what is already visible in the data”. Data from the coming years, the authors write, “will tell us which scenario we are in”. For now, “none of the three can be ruled out”.
(Edited by Tony Rai)
Also Read: ‘This isn’t a marketing stunt. AI could kill all humans.’ Anthropic researcher quits, sounds alarm