Robert Solow's 1987 quip—"You can see the computer age everywhere but in the productivity statistics"—has acquired a new target. Artificial intelligence is demonstrably transforming how knowledge work is performed, yet aggregate productivity growth in most OECD economies remains stubbornly low by historical standards. Either AI is less transformative than claimed, or the way we measure productivity is missing what matters.
Dai (2025) reviews the empirical literature on AI and productivity, finding evidence for effects at multiple levels that do not always align. At the firm level, AI adoption is associated with measurable productivity gains—faster document processing, improved demand forecasting, more efficient resource allocation. At the industry level, the evidence is more mixed, with gains concentrated in sectors that were already data-intensive (finance, technology, advertising) and minimal in sectors that constitute the bulk of economic output (healthcare, education, government, construction). At the macroeconomic level, aggregate total factor productivity growth shows no clear acceleration attributable to AI. The review proposes several explanations for this mismatch: implementation lags (firms need years to reorganize around AI capabilities), redistribution effects (AI may shift value between firms rather than creating new value), and measurement failures (GDP accounting was not designed for an economy where much of the output is free digital services).
Amirov (2025) focuses on the measurement dimension, arguing that traditional national accounting frameworks systematically undercount AI- and data-intensive investment. Because AI models, proprietary datasets, and trained algorithms are intangible assets, they are often expensed rather than capitalized in corporate and national accounts. This means that a firm investing heavily in AI appears to be consuming resources rather than building capital, artificially depressing measured productivity growth. The gap between measured and actual investment may be substantial: Amirov estimates that intangible AI-related investment is growing at rates that, if properly capitalized, would meaningfully narrow the productivity paradox. The implication is not that the paradox is illusory but that our accounting infrastructure has not caught up with the structure of the modern economy.
Li, Liu & Lu (2024), with 14 citations, push the argument further by empirically testing whether digitalization should be treated as a distinct factor of production. Using stochastic frontier analysis (SFA) on Chinese provincial data, they position digitalization as an explicit input in the production function alongside capital and labor, dissecting its elasticity impact on both traditional inputs and total factor productivity. Their key finding is striking: digitalization not only improves output efficiency directly but also enhances the productivity of capital and labor—a complementarity effect that standard growth accounting frameworks miss entirely. When digitalization is omitted from the production function, its contributions are absorbed into the TFP residual, inflating the apparent "unexplained" component of growth. This empirical demonstration from China's digital transformation addresses a genuine measurement challenge: the Solow growth model's residual (total factor productivity) was always a measure of our ignorance, and digitalization—including AI—may be expanding that ignorance faster than the residual can absorb it.
The practical implication for policymakers is sobering: the productivity statistics they rely on to calibrate fiscal and monetary policy may be increasingly disconnected from economic reality. The AI revolution may be real and significant, but its economic signature is being lost in the noise of outdated measurement frameworks.