Fed study finds AI’s productivity boost may be hidden by cheaper output

A study by the Federal Reserve Bank of St. Louis analyzed close to 490,000 earnings-call transcripts from more than 5,000 publicly traded U.S. companies spanning 2000-2025, tagging sentences about productivity and artificial intelligence. By the end of 2025, AI-related commentary accounted for about 15% of all productivity discussion, yet 95% of those remarks projected future benefits rather than reporting realized gains, and overall productivity statistics remain unchanged.

Co-author Serdar Ozkan warned that AI’s ability to make outputs dramatically cheaper can simultaneously erode their value, causing real improvements to disappear from aggregate measures. The paper draws on Robert Solow’s observation that computers took decades to show up in productivity data and likens the current situation to the “productivity paradox” identified by Erik Brynjolfsson. Complementary Fed research from Kansas City and San Francisco finds that early productivity gains are limited to a narrow set of industries and that AI has not yet displaced workers, though firms citing AI positively are increasing R&D and capital spending. The authors conclude that the true impact of AI may only become evident after a prolonged diffusion period, mirroring past general-purpose technologies.