Harvard and Brookings scholars use a four-scenario model and task analysis to show AI has not yet triggered mass layoffs, though 41% of work tasks can already be automated or augmented

The NBER working paper by Elmendorf, Dynan, and Sheiner constructs four scenarios for AI's impact on the economy, ranging from a moderate boost to GDP with little reduction in the number of workers to much faster GDP growth with persistently high unemployment. Rather than offering a single forecast, the work organizes AI's macroeconomic consequences into a set of parallel scenarios and explicitly declines to speculate on their relative likelihood. Based on the authors' economic modeling and scenario construction; the text gives no probability weights, and the co-authors 'cautiously declined to speculate on the relative likelihood of their scenarios.'

The authors use the early-2000s 'China shock' as a comparison: other researchers estimate that disruption cost about 1.5 to 2 million total jobs between 2000 and 2007, while their AI scenario has larger effects, with 3 million people, roughly 2 percent of the labor force, out of work at any given time, and therefore many millions over the next few decades. It places AI's employment shock against the magnitude of a previously studied trade shock, giving the abstract phrase 'large-scale' a comparable scale. The 1.5 to 2 million China-shock figure comes from other researchers cited by the authors; the 3 million figure is a value inside the authors' scenario, not an observed outcome.

Joseph Fuller, in collaboration with Accenture Research, developed an AI model that analyzed work tasks for all job categories and found that as of today 41 percent of all work tasks can be automated or augmented by AI, yet firms are not rushing to automate, with only about one-third of their AI experiments successful. It separates what AI can do from what firms actually do, pointing to a gap between technical feasibility and organizational adoption. Based on an industry collaboration with Accenture Research and a task-analysis model covering all job categories; the roughly one-third success rate comes from Fuller's account, with no sample or statistical detail given in the text.

Raffaella Sadun and Jorge Tamayo at the HBS Digital Reskilling Lab work with firms to retrain workers for AI-augmented roles and then test whether the training works, while Fuller helped one large tech company rewrite entry-level job descriptions so new hires do some software engineering while also getting regular structured introductions to sales and product management. It shifts the discussion from whether layoffs happen to how entry-level jobs and human-capital investment can be redesigned, offering testable training and job-redesign practices. Drawn from lab and firm collaboration practice and a specific company case; the text reports no quantified training outcomes, and the company is not named.

This work speaks to economists, policymakers, and managers focused on AI and the labor market. The four-scenario framework is meant for assessing combinations of AI's effects on GDP and unemployment, not for predicting a specific date; the 41 percent task-automation estimate applies at the task level across all job categories today, not to substitution that has already occurred; the retraining and job-redesign practices apply to firm settings willing to invest in human capital. Elmendorf's proposals, including sweeping changes to the tax scheme, safety-net programs, and unemployment insurance, subsidizing private employment, encouraging shorter work weeks, expanding retraining, and a wage-insurance program paying laid-off workers part of the difference between old and new wages, are framed as preparation for changes that might happen but are not certain.

Readers should still watch several things: the relative likelihood of the four scenarios is deliberately left open, so the text cannot tell which outcome is more probable; the 41 percent task-automation figure is a task-level technical estimate, and organizational adoption still stands between it and actual job loss; the roughly one-third success rate of firms' AI experiments and whether retraining works are not quantified in the text; and Elmendorf himself stresses that what has been seen so far in the labor market has 'very little predictive power' for five, ten, or fifteen years out. In addition, this is a full-text read but does not include figures or raw data tables, so checking specific numbers and model settings requires returning to the original NBER working paper.