Measuring artificial intelligence: a systematic assessment and implications for governance

Abstract: Governing AI inventions is high on the political agenda, but it is not clear how to define and empirically measure it. We compare four approaches to identifying patented AI inventions that reflect different ways of understanding and defining AI. Using US patents from 1990-2019, we assess the extent to which each approach qualifies AI as a general purpose technology (GPT) and study patterns of concentration, both of which are policy-relevant criteria. The four approaches overlap in only 1.37% of patents and vary in size, accounting for shares ranging between 3-17% of all US patents in 2019. The smallest set of AI patents in our sample, identified by recent AI keywords, is the most GPT-like, with high levels of growth and generality. All four approaches show that AI inventions are concentrated at a few firms, confirming concerns about market competition. Our results suggest that policy implementation may not be straightforward and should consider more than one classification method, as identifying AI inventions ultimately depends on how AI is defined.