Analysis · Technology
AI Patents and Research Papers Told Different Productivity Stories
By AWEI · AI-compiled · Published · Analysis prepared · 1 source · doi.org
An AI progress dashboard can change its apparent message depending on what it counts. In the same international study, accumulated AI patents showed little relationship with productivity growth, while accumulated scientific publications showed positive associations in several models. Neither measure directly tracked workplace adoption.
Aleksandra Parteka and Aleksandra Kordalska reported this contrast in a working paper revised in September 2022. They matched AI patent data covering 1985–2016 and publication data covering 1998–2017 with national productivity records, comparing OECD economies with a broader sample including non-OECD economies. Their main outcome was growth in output per hour worked.
The researchers examined annual activity and estimated accumulated knowledge. To build those knowledge stocks, they combined new patents or papers with earlier records, assuming 15% annual depreciation. This matters: a stock represents knowledge retained across years, rather than just the latest research output. It remains a constructed indicator, not a direct measurement of usable AI capability.
Their models accounted for capital deepening, earlier productivity levels, and country and time effects. They also tested controls for general innovation and other national characteristics. Instrumental-variable estimates attempted to address reverse causality. The weak patent relationship broadly persisted across checks; some publication associations weakened after accounting for overall scientific activity. The authors suggest publications may capture a wider range of AI innovation, but the contrast does not establish why the results differ.
These are observational national estimates, with causal interpretation dependent on statistical assumptions. A statistically insignificant patent coefficient does not prove zero gains. Patents and papers capture knowledge production; they cannot establish whether a particular organization deployed AI successfully. The historical observation periods also limit what the findings can say about subsequent technologies.
For an AI value assessment, the useful distinction is between producing knowledge, using it, and increasing output. This study examined the first and third at national scale. Its results make innovation counts an insufficient substitute for evidence about deployment and measured productivity.
Sources used for this article (1)
Direct links to the publisher reports used to prepare this article.
- Source 1
- Artificial intelligence and productivity: global evidence from AI patent and bibliometric data Aleksandra Parteka, Aleksandra Kordalska
