Tracking Artificial Intelligence Innovation in Patent Data - A Systematic Comparison of Four Classification Approaches
Kestin Hötte, Taheya Tarannum, Vilhelm Verendel & Lauren Bennett

Artificial Intelligence (AI) is currently making rapid advances in technological fields, but tracking radical novelty is one of the biggest challenges for innovation scholars. Patent-based approaches might be one of the best data sources that we have on long-term technological evolution. Yet, a study of the literature reveals that there is not yet an established consensus about the definition of AI. This suggests the need to understand different available options and the choice of the classification scheme.

In this paper, we examine artificial intelligence innovations in US patents granted between 1990 and 2019 using four widely-used classification approaches based on (i) keyword searches, (ii) scientific citations, (iii) the World Intellectual Property Organization (WIPO) method, and (iv) the United States Patent and Trademark Office (USPTO) classification. We compare and evaluate the groups of AI patents by their time trends, overlap, reliance on public support, industry links, and technological and scientific base.

Strikingly, only 1.36\% of the total sample of more than 735,000 candidate AI patents are simultaneously identified by all four approaches; the overlaps are only at or less than 20\% for pair-wise comparisons of the different definitions. The approaches differ much in volume, classifying between 2-14\% of all patents granted in 2019 as AI. Depending on which AI classification approach is chosen, it can suggest different patterns of growth, reliance on public and governmental support. Yet, the four samples of AI patents also have qualitative similarities in their technological and scientific characteristics as well as industry links. We highlight approach-specific differences in the suggestive path of growth of AI innovations.

For example, the science and USPTO samples capture more AI innovations at an early phase of the technology cycle, while patents identified by WIPO and keyword method reveal an accelerated growth in later years. The last two approaches also show a larger representation of patents related to machinery manufacturing. Taken together, research aiming to measure AI innovation need to be aware what the choice of AI classification method could mean for their research questions.
Date: 28 October 2021, 15:00 (Thursday, 3rd week, Michaelmas 2021)
Venue: Virtual Seminar via Zoom (please register to attend)
Speaker: Various Speakers
Organising department: Institute for New Economic Thinking
Organiser: Susan Mousley (INET Oxford Admin Team)
Organiser contact email address: events@inet.ox.ac.uk
Part of: INET Oxford Researcher Seminars
Booking required?: Required
Booking url: https://www.inet.ox.ac.uk/events/tracking-artificial-intelligence-innovation-in-patent-data-a-systematic-comparison-of-four-classification-approaches-kestin-hötte-taheya-tarannum-vilhelm-verendel-laur
Audience: Public
Editor: Susan Mousley