Total factor productivity dynamics and the artificial intelligence paradox: Evidence from long-memory analysis

This paper investigates the artificial intelligence (AI) productivity paradox using total factor productivity (TFP) from 1890 to 2022 and fractional integration and long-memory econometric methods. We find that total factor productivity gains from AI investments may be delayed and diffuse nonlinearly, following long-memory patterns similar to those of previous technological revolutions, resulting in a paradox of long lags, not a lack of innovation. The average TFP growth rate (0.54%) in the AI era is the lowest of any post-war technological wave, with profoundly contradictory persistence measures from the GPH (d=1.730) and Local Whittle (d=0.133) estimators, reflecting fundamental uncertainty about the actual productivity of AI. We observe that the GPH estimator is consistent with the "J-curve" hypothesis of temporary slowdown before long-term gains. In contrast, the Local Whittle estimator suggests productivity effects that may be fleeting and easily commoditized. Cross-country heterogeneity in AI persistence patterns points to the role of local institutions, policies, and complementary investments in mediating the macroeconomic impact of AI. These results imply that the full productivity benefits of AI may be realized over very long-run horizons, providing policymakers and investors with necessary guidance on the timing and nature of the AI revolution.