Analysis · Technology
AI Mentions, Energy Use and the Missing Deployment Evidence
By AWEI · AI-compiled · Published · Analysis prepared · 1 source · doi.org
Counting AI terms in an annual report can reveal a statistical pattern without showing what happens on a factory floor. In Huipeng Li’s 2025 study in Scientific Reports, that distinction matters: companies with higher AI keyword scores also tended to have higher green productivity, lower total energy consumption and more green patent applications.
The study examined Chinese listed companies using repeated observations over time. Li constructed a dictionary from an initial manually compiled lexicon, expanded it with a model that identifies related words, then removed irrelevant terms and merged synonyms. Annual report keyword counts, transformed logarithmically after adding one, served as the measure of AI adoption.
The main outcome, green total factor productivity, combines production efficiency with environmental considerations, including resource use and pollution. Fixed-effects regressions included controls for leverage, liquidity and asset composition. The positive association persisted under alternative estimation and after excluding observations from 2007–2008.
The separate energy and patent analyses make the result more interesting, but also demand care. Lower total energy consumption does not by itself establish improved energy efficiency: that requires relating energy use to output. More green patent applications indicate inventive activity, without establishing implementation. Both findings are consistent with the proposed green transformation mechanism; neither independently demonstrates that deployed AI caused operational improvements.
Causal interpretation remains conditional. An additional analysis used firms’ initial AI exposure combined with service-sector growth outside their city, assuming that this growth affected green productivity only through AI. That assumption is not guaranteed by excluding the home city. The paper also reports conflicting study periods: 2007–2020 in the methods and 2007–2022 elsewhere. Its disclosure-based measure cannot establish deployment intensity.
For a business assessing AI value, the useful lesson is to connect each indicator to the claim it can support. Report language, patent applications, energy consumption and productivity describe different things. This study motivates checking their connections in actual operations; it supplies no guaranteed savings estimate for a particular deployment.
Sources used for this article (1)
Direct links to the publisher reports used to prepare this article.
- Source 1
- Artificial intelligence, green transition and green total factor productivity in enterprises Huan Li
