The AI Productivity Paradox: Where’s the Beef?

The Federal Reserve Bank of St. Louis has released a study on the productivity impact of artificial intelligence (AI), reigniting a long-standing debate. Proponents argue that AI’s slow adoption rate is merely a temporary setback, with its benefits far outweighing short-term drawbacks. Detractors counter that the technology has failed to deliver on its promise and that its benefits are being overstated.

At the heart of this debate lies a fundamental question: what exactly do we mean by “productivity” in the context of AI? Is it simply about increasing output, or is there something more nuanced at play? The answer, much like the technology itself, is complex and multifaceted.

The St. Louis Fed study’s findings are telling. By analyzing nearly 490,000 corporate earnings calls, researchers Serdar Ozkan and Aakash Kalyani found that AI-related productivity gains have been largely hypothetical, with executives expressing optimism about future benefits rather than actual results. This mirrors the historical pattern of technological adoption, as economist Robert Solow famously noted: “You can see the computer age everywhere except in the productivity statistics.”

The study highlights the need to rethink our metrics for measuring productivity. Ozkan and Kalyani pointed out that productivity is not a single number but an output of an entire chain of processes. While AI may have accelerated some links in this chain, others remain stubbornly slow.

The concept of “abundance” also comes into play here. When AI makes certain outputs cheaper to produce, their value simultaneously plummets. In other words, the productivity math cancels itself out: gains in one column get erased by falling prices in another. This is not a new phenomenon; we’ve seen it before with electrification and other technological revolutions.

Economists Erik Brynjolfsson and Nicholas Sullivan have described this scenario as the “productivity paradox,” where new technologies fail to deliver expected gains in productivity. However, this may be less a paradox than a reflection of our own limited understanding of what it means to be productive. Kalyani’s comment that “the application that ends up mattering gets discovered through trial and error” hints at the uncertainty inherent in technological adoption.

The study also highlights the limitations of our current metrics. We’re measuring productivity gains based on narrow definitions of output, rather than considering the broader ecosystem within which these outputs exist. The San Francisco Fed study’s findings on AI-positive firms investing more in R&D and capital expenditures are telling but also raise questions about our expectations for technological adoption.

Brynjolfsson’s concept of a “modern sequel” to the original productivity paradox raises important questions about our assumptions regarding technological impact. Have we been too hasty in assuming that AI will deliver on its promise of boosting aggregate productivity? Or are we simply being naive about the time it takes for new technologies to reorganize entire industries?

The slow diffusion of AI adoption across regions, occupations, and firms is a sobering reminder of how difficult it is for new technologies to reorganize entire industries. We should be cautious about assuming that AI will deliver on its promise in the short term.

Despite AI’s ability to accelerate some links in the productivity chain, bottlenecks remain stubbornly slow. The example of Google’s early struggles with Googolplex highlights the uncertainty inherent in technological adoption – and the need for continued investment and experimentation.

The abundance argument suggests that it may be a matter of time before AI finally delivers on its promise of boosting aggregate productivity. However, this raises the prospect that some benefits may remain invisible to our metrics. The honest answer is that nobody knows in advance what will finally prove AI’s productivity gains have arrived – but with continued investment and experimentation, we may yet uncover the secrets of this enigmatic technology.