AI Won't Transform The Economy As Quickly As We Think
While advocates of each technological revolution often make it seem like their technology is unlike any that has gone before, the reality is that the adoption of new technologies tends to follow fairly. . .
While advocates of each technological revolution often make it seem like their technology is unlike any that has gone before, the reality is that the adoption of new technologies tends to follow fairly familiar patterns. The history of automation is written in wheat fields and factory floors. Where once dozens of workers threshed grain by hand, a single combine harvester now does the job. Spot-welding in car plants has passed from human hands to industrial robots. Yet for all the drama of such transformations, a provocative newpaperfrom Stanford University suggests that even the most sophisticated automation, including artificial intelligence, may deliver economic gains more slowly than many expect. The authors’ central finding is both startling and sobering. Using a task-based model calibrated to historical data, they estimate that by 2040, GDP per person will be just 4% higher than it would have been without AI-driven growth acceleration. Even by 2060, the gain reaches only 19%. This is not because automation fails to boost productivity—quite the opposite. Rather, it reflects a fundamental feature of how modern economies work: the stubborn persistence of what economists call “weak links. ”The weak-link economyThe intuition is straightforward. Imagine production as a chain where each link represents a different task. Some tasks may be performed with extraordinary efficiency by machines that improve at 5% annually. Others remain the province of humans, whose productivity creeps forward at perhaps 0. 5% per year. When tasks are complements rather than substitutes—when you need all of them to produce output—the economy is constrained by its weakest links. This helps explain a puzzle the authors uncover in historical data. Between 1950 and 2023, capital productivity in America’s private business sector grew at least five percentage points faster annually than labour productivity. The main benefit of automation was not the switch itself, but rather the ongoing rapid improvement in machines compared with the glacial progress of human workers. Yet if automation had been frozen at 1950 levels, the authors calculate that essentially all subsequent productivity growth would have been eliminated. The economy’s advance depended on steadily strengthening more of its weak links. The Baumol bottleneckThis phenomenon was famously chronicled by the economist William Baumol, who observed that some activities resist productivity improvements. A string quartet takes as long to play a Mozart piece today as it did in 1791. The same logic applies more broadly: elderly care, complex negotiations, and creative problem-solving. If AI cannot fully automate these tasks, they become binding constraints on growth. The Stanford researchers modelled three different scenarios over the long term. In one, a small fraction of tasks, typically around 5%, can never be automated. Here, the capital share of GDP eventually falls to zero and growth slows to the rate at which humans improve at the remaining tasks: Baumol’s cost disease writ large. In another scenario, every task is eventually automated, and growth explodes. The intermediate case, where automation approaches but never quite reaches 100%, also features explosive growth, but one that unfolds with surprising lethargy. Mind the gapEven the intermediate scenario suggests annual growth rates could reach 5% within 75 years. Yet the authors’ models show remarkably similar paths across all three scenarios for the next several decades. This matters for policy. If AI’s transformative effects arrive slowly, societies will have more time to adapt labor markets, retrain workers, and address inequality. It also suggests that the current AI boom, while significant, may not demand the immediate, dramatic policy responses some advocate. The analysis rests on several assumptions, including that machines will continue improving much faster than humans at the tasks they perform. The authors measure automation rates using queries to ChatGPT, which is an approach they acknowledge needs refinement. Yet the paper’s core insight seems robust: in an economy characterised by complementary tasks, even infinite productivity on some tasks delivers only finite gains. A nuanced pictureFor investors and executives, the message is nuanced. AI will likely accelerate economic growth, but the bonanza may take longer to arrive than current euphoria suggests. A 25% boost to research productivity, the authors calculate, could bring forward the gains substantially—raising output 22% above trend after 20 years rather than 4%. This highlights the importance not just of deploying AI, but of directing it toward multiplying human capabilities across a broad range of tasks. The future, it seems, will be automated. But weak links ensure it will not arrive overnight. Those hoping to repair the computer, negotiate the business deal, or teach the kindergarten class need not worry about obsolescence just yet. The explosion, when it comes, will be a slow one.