« CSET Foretell Blog

The U.S. AI Industry’s Demand for AI-Skilled Researchers Is Increasing, but Not Its AI Publications

Published
Aug 17, 2020 02:05PM UTC
Author
Michael Page and Roxanne Heston

This post is part of our Metric Analysis series, where we explore historic data related to the metrics (questions) we are forecasting on Foretell.


From 2013 to 2019, the rate of AI-relevant publications at prominent tech companies—Apple, Amazon, Facebook, Google, and Microsoft—was relatively constant while the rate of job postings for machine learning-skilled researchers at those companies increased more than seven times.1 These trends are displayed in Figure 1. A possible explanation is that a progressively smaller percentage of these companies’ research is amenable to publication, or that each publication requires more research effort. An alternative explanation is that these companies are electing to publish a smaller percentage of their research. Our analysis explores this trend but not its cause.

Figure 1. This chart is based on Dimensions (publications) and Burning Glass (job postings) data. Using a predictive model trained on arXiv data, we classified publications as AI/ML-relevant or not. A publication is deemed relevant if it’s posted on arXiv under any of these categories: artificial intelligence, machine learning, computer vision, computation and language, multiagent systems, or robotics. To read more about this method, see “Identifying the Development and Application of Artificial Intelligence in Scientific Text.” We included a Burning Glass job posting if it included machine learning among the required skills and categorized the occupation as “Researcher / Research Associate.” Data pulled: August 13, 2020.


Measuring AI-related research activity is less straightforward than measuring AI publications. We consider two different metrics of research activity. One way to measure AI-related research activity at an organization is by the number of machine learning-skilled employees in research roles. As displayed in Figure 2, from June 2019 to June 2020, the number of machine learning-skilled researchers increased at each of the five companies. Across all five companies combined, it increased by 27 percent.  


Figure 2. This chart is based on Linkedin Talent Insights data. We counted people whose listed skills included machine learning and whose job function Linkedin classified as “Research.” Data pulled: August 13, 2020.


While employment data suggests the rate of AI-related research activity at these companies is increasing, this inference is based on only two data points. Because we have only two employee-based data points, we can use the employees/postings metric to make comparisons across companies but our ability to make comparisons over time is limited. An alternative measure of AI-related research activity that allows for comparisons over a greater time period is job postings for machine learning-skilled researchers. As shown in Figure 3, the number of machine learning-skilled researchers at a company correlates with job postings for machine learning-skilled research roles. Because of this correlation, we feel comfortable using research job postings as a proxy measure—albeit an imperfect one—of research activity.  

Figure 3. This chart is based on Linkedin Talent Insights (employee) and Burning Glass (job postings) data. It includes the top 20 for-profit companies by machine learning-skilled research job postings over two years, comparing the quantity of job postings to the quantity of machine learning-skilled researchers as of August 12, 2020.


As seen in Figure 4, the publications/postings metric shows a consistent downward trend, suggesting either a change in the type of research occurring within these companies or the percentage of research that’s published. Although Figure 2 shows disproportionate growth in researchers at Amazon, even if Amazon is removed from the dataset the publications/posting metric shows a similar trend. This metric supports a forecast question on Foretell and informs a world forecast "Three Possible 2025 Worlds that Should Inform Policy Today." You can forecast it here.  

Figure 4. This chart is based on Dimensions (publications) and Burning Glass (job postings) data. The Y-axis is the number of AI publications in the calendar year divided by the number of job postings for research roles requiring machine learning skills in a calendar year. See Figure 1 for how we identified AI publications in Dimensions. For Burning Glass job postings, we included a job posting if Burning Glass included machine learning among the required skills and categorized the occupation as “Researcher / Research Associate.” Data pulled: August 13, 2020.



Footnotes

1.  During this period, AI publications as defined in Figure 1 represented 30 percent of all publications for these companies, and job postings for machine learning-skilled researchers as defined in Figure 1 represented 53 percent of all job postings for researchers at these companies.



Authors: Michael Page and Roxanne Heston, CSET