Measuring artificial intelligence: a systematic assessment and implications for governance

Abstract Governing artificial intelligence (AI) inventions is a major policy concern. Yet, definitions and measurement approaches remain contested. We compare four patent-based definitions reflecting distinct understandings of AI. Using US patents (1990–2019), we assess the degree to which each approach describes AI as a general-purpose technology (GPT) and examine patent concentration by a few dominant firms. We find that between 3% and 17% of all US patents in 2019 are classified by at least one of the approaches as AI patents. Yet, only 1.4% of all AI patents are simultaneously identified by all four approaches. All approaches are consistent with AI having GPT characteristics, with the Keyword-based patents exhibiting the highest growth and generality. GPT indicates public good characteristics, which could be used to justify public support. Across methods, AI patents are concentrated among a few firms, highlighting market power and regulatory challenges. The wide variation in the subsets and characteristics of AI patents identified by these approaches suggests that currently multiple classification methods should be considered to formulate robust, inclusive, and effective analyses for AI governance.