AI Could Eliminate 20 Million US Jobs While Adding $2 Trillion

A unified framework for AI's economic impact

Economists have produced wildly contradictory forecasts for artificial intelligence's effect on the economy. Daron Acemoglu estimates cumulative productivity gains under one percent over a decade. Goldman Sachs analysts project $7 trillion in global output gains. McKinsey speaks of trillions in annual value. When expert predictions diverge this dramatically, the problem typically lies not in the data but in the models themselves.

Benjamin Verschuere and Angus Cameron argue in a new working paper that this dispersion stems from two fundamental gaps. First, economists build estimates for growth, employment, and prices in isolation, with no framework to reconcile them. Second, models fixate on AI's current technical capabilities rather than measuring how quickly it spreads through the economy and how much of any given job it can ultimately reach.

The researchers developed a unified framework that addresses both gaps, producing internally consistent forecasts across multiple economic variables. Their model projects roughly $2 trillion in long-run US output gains, the elimination of approximately 20 million American jobs, and a seven to eight percent permanent reduction in the overall price level.

Why it matters

This research provides policymakers with the first comprehensive view of AI's economic trade-offs within a single analytical structure. Unlike fragmented forecasts that predict large productivity gains without accounting for displaced wages or price effects, this framework forces every projection to add up. The finding that existing fiscal channels can support displaced workers while consumers benefit from lower prices suggests the transition is manageable—but only if labor markets remain flexible and AI markets stay competitive. The alternative is productivity gains that concentrate with capital rather than spreading through the economy.

Measuring what AI can actually automate

The framework starts at the task level. The researchers developed a metric called "workflow completeness" that estimates the share of a job's tasks AI can complete end to end. Combined with Amdahl's Law—which holds that any process is limited by the component that cannot be accelerated—this approach recognizes that if one-fifth of a job must remain human, the entire workflow moves no faster than its human component.

Aggregating from tasks to industries, the model finds AI can automate about 12 percent of construction, roughly 68 percent of information technology and finance, and 75 percent of customer service. These structural limits don't rise as AI models improve because they're constrained by physical, regulatory, and relationship-bound work that better software cannot touch.

Four rates that govern diffusion

The researchers decompose AI's progress into four distinct rates. The innovation rate measures how quickly AI becomes technically capable of performing a task. The deployment rate tracks how fast firms place tools in workers' hands. The adoption rate—the most critical for productivity—measures how quickly organizations redesign work to capture real gains. The displacement rate tracks the conversion of productivity gains into actual headcount reduction.

Deployment currently runs well ahead of adoption. The tools are widespread, but most organizations haven't rebuilt workflows around them, leaving productivity gains largely uncaptured. Displacement lags even further behind and has barely begun, which explains why AI feels pervasive while aggregate statistics remain quiet.

Evidence AI has already reshaped exposed sectors

The most AI-exposed US sectors are already 11 percentage points more productive than before OpenAI launched its first widely available model in 2022. AI exposure explains roughly 70 percent of cross-sector variation in productivity gains, representing about $642 billion in output above the pre-2022 trend.

Jobs in the most exposed sectors run approximately 2.9 million below their pre-pandemic trajectory, driven mainly by suppressed hiring rather than layoffs. Wages in these sectors have grown slower than productivity, with the gap scaling almost exactly with how many tasks can be automated.

The policy challenge

The central implication is that the transition is positive on balance. The $2 trillion in productivity gains more than offset the cost of displaced workers while delivering lower prices. The resulting fiscal capacity is sufficient to support the displaced through existing channels.

The challenge lies in where gains land. Currently, the surplus is captured by companies selling AI products and firms buying them, not reaching consumers. Competition erodes this concentration—as rival firms adopt the same tools, productivity gains convert from profits to lower prices.

The policy task requires keeping AI markets competitive and lowering barriers that prevent displaced workers from finding new employment. If markets fail to absorb displaced workers, AI will reduce aggregate income and demand. The researchers emphasize that monetary policy cannot address this structural challenge; instead, fiscal channels must support workers while competition ensures consumer welfare benefits materialize.

The full working paper and a condensed version were published by researchers Benjamin Verschuere and Angus Cameron, who are employed by Liminal Capital LLC, and first reported by ProMarket.