Fed study: AI productivity gains vanish as output gets cheaper

Economists Serdar Ozkan and Aakash Kalyani, along with research associate Nicholas Sullivan from the Federal Reserve Bank of St. Louis, conducted a sweeping analysis of corporate earnings to understand AI's actual economic impact. They scanned roughly 490,000 earnings call transcripts from 5,198 publicly traded U.S. firms between 2000 and 2025, using an AI model to tag sentences about productivity and AI. The results reveal a striking temporal gap: the share of productivity commentary tied to AI rose from near zero before ChatGPT's late-2022 debut to roughly 15% of all productivity discussion by the end of 2025. Yet when executives describe AI's effect, approximately 95% of AI-related productivity sentences describe gains they expect in the future, not gains already achieved. This ratio has remained steady since 2023. When executives do speak of productivity effects, they are almost uniformly bullish—95% describe productivity as rising, compared with only 75% for non-AI productivity commentary.

The absence of measured productivity gains is not new. Official aggregate data have shown no meaningful productivity bump once capital investment is accounted for over the past three years. Ozkan said he wasn't surprised by the future-tense findings, invoking economist Robert Solow's famous 1980s observation that 'you can see the computer age everywhere except but in the productivity statistics.' Ozkan drew a direct historical parallel to electrification, which took 'several decades' to reorganize factories, retrain workers, and change workflows before its productivity payoff showed up in the data. This mirrors a pattern identified by Stanford economist Erik Brynjolfsson in a 1993 MIT paper, which he termed the 'productivity paradox.' Kalyani, who has separately studied diffusion patterns across general-purpose technologies, noted that the profession has 'largely reached consensus on this point after the initial post-ChatGPT excitement faded.' He explained that 'the aggregate gains will be in the future, whereas what you see right now is a lot of investment and a lot of excitement and optimism for the future.' Technology diffusion across regions, occupations, and firms is 'extremely slow,' typically unfolding over 20 to 30 years, making a three- to five-year payoff 'a huge change' from historical precedent. Computers, per Solow's observation, didn't show up in productivity data until the late 1990s and early 2000s.

Yet researchers uncovered a mechanism that may explain why measured gains could remain invisible even if real economic improvements are occurring. Ozkan articulated the core insight: 'Some things are going to become more abundant. That means they're also going to become probably less valuable.' When AI makes output radically cheaper to produce—marketing materials, animations, news stories—that output simultaneously becomes less valuable because it is no abundant. The productivity math cancels itself: gains in one column get erased by falling prices in another. A task gets easier, output gets cheaper, and somewhere in that trade a real gain disappears from the statistics without showing up as a loss. A second constraint compounds this invisibility: bottlenecks that AI cannot dissolve. However fast AI accelerates research, drafting, or analysis, two people still need to schedule and show up to a meeting at the same speed as four years ago. Productivity is not a single number but the output of an entire chain, and AI has only sped up some of the links.

Crucially, the St. Louis Fed team emphasizes this is not empty corporate talk. Kalyani stated the researchers 'trust what people do, not what they say,' noting that firms discussing AI positively have also increased R&D, capital expenditures, and investment—a correlation that didn't exist when the team first studied AI mentions in an earlier 2024 post skeptically titled 'AI Hype or Reality?' but has since strengthened. A related San Francisco Fed study found AI-positive firms saw substantially higher investment and R&D growth by 2025 than other public companies, concentrated among the largest technology firms building AI infrastructure. This pattern aligns with other recent Federal Reserve research showing the productivity pickup in official data is 'not yet broad-based,' with a small set of industries accounting for most gains even as AI adoption spreads. Fed Chair Kevin Warsh told Congress in July that AI 'hasn't displaced workers' so far and has made them 'a bit more productive,' but cautioned that 'the long term can be quite far out.' Previous St. Louis Fed research estimated generative AI represented only a 1.1% increase in productivity by late 2024 relative to 2022—modest against the 2.3% and 1.6% overall productivity growth the economy posted in 2024 and 2023 respectively. When asked what signal would finally prove AI's productivity gains had arrived, Kalyani gave an honest answer: 'Nobody knows in advance.' The application that ends up mattering gets discovered through trial and error, spreading firm by firm and worker by worker, in a process that looks almost random even as it adds up to something real. He offered the example of Google Maps and the taxi medallion: nobody predicted which navigation app would end taxi monopolies, yet the pattern repeats with every general-purpose technology—winners and losers get decided by a chaotic, decentralized scramble that resists prediction, even when the technology's eventual importance is obvious in hindsight.