Distance from innovation is a barrier to the adoption of artificial intelligence

The extent to which geographic distance is a barrier to technological knowledge transfer is of interest to governments of countries distant from centres of knowledge creation or technology production; to entrepreneurs deciding where to locate a new firm that will need to remain abreast of technological developments; and to national or local policymakers seeking to influence the decisions of such entrepreneurs. These agents may value knowledge transfer as an input to further knowledge creation, or as a prerequisite for the adoption of new technology practices.

Distance may seem unimportant in the face of technological progress including the telephone, modern means of transportation, email, texting, the worldwide web, and video conferencing. Yet, several studies have found that distance is a barrier to the diffusion of inventive activity and to the cross-country diffusion of technology adoption, and prior work has also shown that US state borders are barriers to citations of patents. ¹ Cross-location collaboration and citing of academic papers and patents have been increased by shorter travel times, other papers have found. ²  

Whether distance constitutes a barrier between innovation and technology adoption has been little studied, however. Distance could be a barrier to adoption by companies already operating in distant commuting zones, or to the establishment in distant commuting zones of companies anticipating adoption. Agha and Molitor (2018) show that doctors are more likely to adopt a new cancer drug if the clinical trials were held in the same region. Bloom et al. (2021) consider a group of 29 ‘disruptive’ technologies, including artificial intelligence (AI), showing they emerge through patents in concentrated ‘pioneer locations’ before spreading geographically as measured by convergence across locations in the share of job advertisements involving the technology group.

In new research (Hunt et al. 2024), we examine the geography of US firms’ adaptation and adoption of AI in response to AI innovation, considering distance explicitly and distinguishing between innovation disseminated through patents and through scientific publications. We choose to examine AI in part because the rapid growth in AI research papers and patents began only recently, allowing an examination of its geographic diffusion from early in the process. It is also of particular interest because it is potentially important for future economic growth. Because AI is still immature, with few off-the-shelf applications yet available, we seek evidence for the effect of distance on the adaptation of AI to a new environment, such as a new industry, in addition to the effect on adoption.

To measure innovation, we create a dataset of AI publications, using Microsoft Academic Graph (MAG) to count journal articles, conference proceedings and patents identified in MAG as relevant to ‘deep learning’. We divide the US into 741 commuting zones, and designate as AI innovation hotspots those commuting zones whose cumulative AI publications before our study period were over a certain threshold. We measure AI adaptation or adoption using job vacancy information from US online job advertisements scraped by Burning Glass Technologies (renamed Lightcast) from 2007-2019, again aggregating to the commuting zone level. With this panel of commuting zones in hand, we investigate whether online vacancies for jobs requiring AI skills grow more slowly in US locations farther from pre-2007 AI innovation hotspots.

Figure 1 displays the commuting zones which are hotspots when the AI publication (patent plus patent) threshold is 1,000. We find that over seven years, a commuting zone which is an additional 200 km (125 miles) from the closest such AI hotspot has 17% lower growth in AI jobs’ share of vacancies. The effect of distance grows more negative as the number of publications in the hotspot rises and is driven by distance from AI papers rather than AI patents. Distance reduces growth in AI research jobs as well as in jobs adapting AI to new industries, as evidenced by strong effects for computer and mathematical researchers, developers of software applications, and the finance and insurance industry. Distance from an AI hotspot has much less impact on the adoption of AI, such as the use of image processing.

Figure 1 Innovation hotspots’ AI publications through 2006

Note: The Arlington, VA, commuting zone includes the Washington, D.C. area.

Twenty percent of the overall distance effect is explained by the presence of state borders between some commuting zones and their closest hotspot. This could reflect state borders’ impeding migration (Wilson forthcoming) and thus flows of tacit knowledge. Distance does not capture the difficulty of in-person or remote collaboration, since travel time has no effect conditional on distance. Nor does distance capture knowledge and personnel flows within multi-establishment firms hiring in computer occupations.

The results are consistent with the related literature finding that distance is a barrier to knowledge transfer, but inconsistent with the hypothesis that distance reduces AI job growth by making in-person or remote collaboration difficult. The contrast between the effects of the (highly correlated) distances to AI papers and to AI patents suggests that studies focusing on spillover effects or other geographic aspects of AI patents alone may be mistaking the impact of scientific papers for an effect of patents.

Source: cepr.org