Content uploaded by Astrid Krenz
Author content
All content in this area was uploaded by Astrid Krenz on Oct 23, 2022
Content may be subject to copyright.
Comparing New Measures of Tech Talent: Global AI,
Digital Infrastructure, and Innovation∗
Richard Dickens†
Astrid Krenz‡
Jacqueline O’Reilly§
Abstract. International comparisons of the development of AI have increasingly been made
with the use of composite indexes. These aim to identify countries that are at the forefront of AI
and those lagging behind. In this paper, we focus in particular on the talent component related
to AI. We analyse a new dataset based on the Global Artificial Intelligence Index and compare
new measures of tech talent workers with traditional education-based measures. We provide an
overview of different measures of AI, explain how tech talent working in AI, Machine Learning,
and Data Science is measured, and which data are utilised. We show that several traditional
measures are well approximated by the new measures. Countries positioning as forerunners or
laggards in terms of new tech talent workers are detected. Our analysis establishes a link between
governments’ R&D and AI-related expenditures or targets and the number of researchers and
new tech talent workers in a country, as well as between talent, the digital infrastructure, and
innovative activities.
Keywords: Artificial Intelligence, AI, Talent, Skills, Governmental Investments, R&D, Digital
Infrastructure, Innovation.
JEL: J23, J24.
∗This work was funded by the ESRC Digital Futures at Work Research Centre (Digit) under grant number
ES/S012532/1, which is gratefully acknowledged. We would like to thank William Hunt for valuable comments
on an earlier version of this paper. The Stata Code is available upon request. Declarations of interest: none.
†University of Sussex, Digital Futures at Work Research Centre, Jubilee Building, Brighton, Falmer BN1 9SN,
United Kingdom; email: r.f.dickens@sussex.ac.uk.
‡University of Sussex, Digital Futures at Work Research Centre, Jubilee Building, Brighton, Falmer BN1 9SN,
United Kingdom; email: a.m.krenz@sussex.ac.uk.
§University of Sussex, Digital Futures at Work Research Centre, Jubilee Building, Brighton, Falmer BN1 9SN,
United Kingdom; email: j.o-reilly@sussex.ac.uk.
1. Introduction
Artificial Intelligence (AI) is on the rise and it is shaping the workplace and society. The
number of AI start-ups in the USA increased 20-fold from 2000 to 2019; students enrolment into
AI courses increased 16-fold internationally between 2010 and 2019; and the number of AI-related
publications rose from 1,000 in 2010 to over 20,000 in 2019 with the most popular topics being
’machine learning’, ’computer vision’, and ’natural language processing’, respectively (Perrault et
al., 2019; Shoham et al., 2018). The development of AI and the connection of human beings to a
digital network are seen among the most important unique events in human history according to
Brynjolfsson and McAfee (2014), and we are witnessing an exponential development of computer
technologies, digital data, and innovation.
AI can be defined as ”understanding and building intelligent entities/ machines that can
compute how to act effectively and safely in a wide variety of novel situations” (Russel and
Norvig, 2022) or in other words ”AI seeks to make computers do the sorts of things that minds
can do” (Boden, 2016). Early AI systems/ machines were programmed to demonstrate a variety
of components of intelligence, encompassing perception and understanding our environment,
machine learning, problem solving and planning, reasoning, and natural language understanding
(Wooldridge, 2018). In the modern era tasks that are being digitalised through algorithms and
software involve for example image and speech recognition or machine translation and they are
executed by robots, intelligence systems, virtual assistants, and autonomous vehicles, to name
just a few emerging technologies.1
Several institutions and consulting companies report about AI-related performance and progress
(e.g. McKinsey (2017, 2018), OECD (2022a), Zhang et al. (2022), for the European Commission
see Lopez Cobo et al. (2021), or the State of AI report by Banaich and Hogarth (2020)). They
visualise and describe the development of different indicators such as AI tools and technical per-
formance regarding speech and image recognition, AI publications and conferences, economic
performance regarding AI start-ups or employment in AI branches, or enrolment into AI courses
and education. Aside from the reports, several attempts have been made to develop composite
measures of AI, namely the Stanford HAI AI Vibrancy Index (Zhang et al., 2021, 2022), the
1Vannuccini and Prytkova (2019) provide an analysis on understanding AI as a general purpose technology versus
AI being a large technical system. The differentiation is important as it leads to different policies and strategies.
Neufeind et al. (2018) collect an array of scientific works that focus on the new technologies’ impact for the future
of work.
1
Tortoise Global AI Index (Tortoise Media, 2022) or the European Commission’s AI Watch Index
(Lopez Cobo et al., 2021).
We add to the existing literature by reviewing different methods of measuring AI. An aim
of our analysis is to disentangle how much we can learn from the new composite measures of
AI and its elements. We analyse the Tortoise Global AI Index dataset and compare different
components related to skills and talent: traditional education-based measures alongside new
tech talent measures. We identify forerunning and lagging countries in terms of AI and tech
talent. Moreover, we investigate how R&D expenditures and AI strategies play in, and analyse
the influence of talent and digital infrastructure for innovative activity.
In the past years, production processes were subject to a surge of automation of tasks (through
digitalisation, robots, AI). Automation technologies can increase labour productivity (value
added per worker) which could lead to a higher demand for labour, and will consequently in-
crease employment and wages. However, this explanation alone disregards the fact that many
automation technologies were designed to substitute human labour through cheaper machin-
ery capital, which will lead to a reduction in labour’s share in value added (Karabarbounis
and Neiman, 2014) and may thus not raise wages and employment sufficiently (Acemoglu and
Restrepo, 2019). Acemoglu and Restrepo explain that aside from these arguments, it is ques-
tionable if and how technological progress can uniformly expand to other, new sectors and make
labour uniformly productive in everything; furthermore, the rise in inequality between high- and
low-skilled workers should not be overlooked.2The authors further underline that what consti-
tutes a danger for labour in the future is whether the productivity effects will be high enough
such as to compensate for the displacement effects, and they ask ’Do we go for the wrong or
the right kind of AI’, an AI that is beneficial for all humans (Acemoglu and Restrepo, 2018)?
Agrawal et al. (2016, 2019) point out that the labour market effects of AI might be ambiguous.
They explain that with AI a tool has been developed to ease making predictions. They conclude
that prediction tasks could be substituted by AI/ machines. However, making decisions, which
is complementary to making predictions, could either be substituted or complemented by AI/
2Krenz et al. (2021) investigate how automation affects firms’ reshoring and which effects result for wages,
employment, and inequality. They find that automation-induced reshoring increases the skill-premium between
high- and low-skilled workers, leading to rising inequality. Mueller et al. (2017) find that wage inequality rises
with firm size. They explain that this might be due to larger firms being more likely to automate routine job
tasks. They provide evidence that wages associated with routine jobs declined relative to those associated with
non-routine jobs as firms become larger. Korinek and Stiglitz (2017) analyse how AI may lead to a Pareto
improvement, how inequality is affected, how taxation can be used to compensate for potential losses and through
which channels technological progress may lead to technological unemployment.
2
machines. Frey (2019) explores how technological progress led to rising prosperity, a declining
labor share of income, a rise of inequality and how it affected employment. He outlines the
prospects of AI and mentions challenges that will lie ahead e.g. in terms of distributional tasks.
While studies on AI and its implications for labour markets are yet to come and depend on the
availability of datasets and the measurement of AI, the past literature offers ample evidence on
the effects of automation, which is often measured by robots’ usage. Frey and Osborne (2017)
found evidence that about 47 percent of jobs in the US are at high risk of being automated
(by AI or other technologies). Graetz and Michaels (2018) showed that higher levels of robot
density can lead to a decline in hours worked by low-skilled workers. Susskind (2020) explored
how technological progress affects prosperity as well as inequality, and how it leads to a world
of less work.
Digitalisation has led to a rise of the so called platform economy. Newer research provides
labour market analyses of the gig economy. This literature has provided some novel measure-
ments by the use of tools that relate to big data analyses. Kassi and Lehdonvirta (2018) created
the so called Online Labour Index which is based on the number of projects and tasks of job
vacancies posted on major online gig platforms. They find evidence for a substantial growth
of gig work over time. Adams-Prassl (2020) investigates the gender wage gap based on data
from the online labour platform Amazon Mechanical Turk. Women are found to earn about
20 percent less than men and the gap is found to be driven by differences in working patterns
between women and men. Krenz and Strulik (2022) develop a new macroeconomic theory on
the servant economy and provide evidence for inequality driving the rise of servant employment
and wages. Their model demonstrates a rise in gig work due to increasing technological progress
and inequality.
Policy makers’ concerns with the development and consequences of AI can be seen from the
European Commission’s initiative in 2018 to establish a coordinated plan for member states col-
laboration and encouragement to develop national AI strategies (European Commission, 2018).
This plan was reviewed in 2021 focussing on further aspects which are to attain a global lead-
ership of the EU in human-centred, trustworthy, secure and sustainable AI, and to compete
globally (European Commission, 2021b). The European Commission has released a white paper
3
on AI in 2020 (European Commission, 2020), with the aim of building an eco-system of excel-
lence involving private-public partnerships alongside an eco-system of trust ensuring consumers’
rights and security.
In order to make people able to cope with the rising demands due to the new emerging tech-
nologies, policy makers around the world have been interested to enforce education in Science,
Technology, Engineering and Mathematics (STEM) (see e.g. the ’New Skills Agenda for Europe’
recently proposed by the European Commission (2021a)). The European Centre for the Devel-
opment of Vocational Training (Cedefop) regards it as vital to modernize Europe’s education
and training systems and job markets in order to remedy the digital divide and the digital skill
gaps (Cedefop, 2018). According to Cedefop’s European Skills and Jobs Survey (ESJS) about
85 percent of all EU jobs need at least a basic digital skills level. Skills in STEM have been
found to be important for attaining higher wages as well as STEM or STEM-related occupations
(see e.g. Black et al., 2021).
Based on these policy concerns and the increasing use of composite indexes to measure the
development of AI, we examine how well these indexes capture these trends in AI and in tech tal-
ent, in particular. Given the availability of variables in our dataset, we examine both traditional
education-based measures that relate to students and researchers in STEM-related fields, i.e.
Researchers in STEM, IT graduates, IT undergraduates, Science graduates, Science undergradu-
ates, and STEM graduates. Moreover, we define the term tech talent to denote occupations that
classify work in AI, Data Science, and Machine Learning such as AI Engineers, ML Engineers,
Data Engineers, Data Scientists, as well as measures on AI Meetup members, GitHub Stars,
GitHub commits, and Kaggle Grandmasters.3
Our analysis uses for the first time a new dataset that is based on the Global Artificial
Intelligence Index from Tortoise Media (Tortoise Media, 2022). Their AI index is constructed
with expert knowledge from academia, industry and the government, such as the Alan Turing
Institute or Best Practice AI. The dataset contains a rich set of variables on economic factors,
development, talent and education, infrastructure, research and innovation.
Based on these data, our analyses show that several new tech talent measures are in line with
traditional measures on IT, STEM or Science students. In particular, measures of AI Engineers
3Harrigan et al. (2020) constructed a new measure of techie intensity which is based on jobs in STEM areas.
Based on French firm data the authors show that techie intensity led to polarization. Moreover, the authors can
show that firms that had more techies grew faster.
4
and of Data Scientists are displaying a strong connection to measures relating to the number of
Researchers in STEM and to the number of IT graduates.
For Europe, we further identify countries that belong to the forerunners (Western European
and Scandinavian countries) or laggards (South and Eastern European countries) in terms of
global AI, tech talent, and skill- based measures. We reveal a link between governments’ R&D
expenditures and AI-related targets and the number of researchers as well as new tech talent
workers. Moreover, we find a link between talent, the availability of digital infrastructure, and
innovative activities.
The structure of the paper is as follows. Section 2 reviews the literature and explains the
different ways of AI measurement. Datasets are discussed in section 3. Section 4 provides an
empirical analysis, comparing new tech talent measures versus traditional education-based mea-
sures, finding forerunners and laggards in terms of those measures, and analysing governmental
expenditures, AI targets, and the link between talent, digital infrastructure and innovation.
Our analyses reveal that AI measurement in the current literature reveals some new evidence,
however it also correlates with development of STEM-related figures. The last section concludes.
2. Measurement of Artificial Intelligence
There have been a growing number of organisations attempting to measure and compare
countries’ performance along a range of AI measures over time. In part, this is motivated by
identifying new indicators that could allow us to measure how this is developing. There is also
an interest in knowing which countries are performing particularly well and which are lagging
behind and why this is the case. We can differentiate between measurement at the macro-level,
which encompasses country studies and comparisons, and measurement at the micro-level, which
deals with the level of the firm, individual, sector or occupation.
As regards the measurement at the macro-level, the Stanford Institute for Human-Centered
AI (HAI) have developed the annual Global AI Vibrancy Tool since 2017. This tool permits an
interactive visualization comparing 29 countries across 23 indicators. The indicators focus on
research and development, the economy, and inclusion. More precisely, they comprise measures
such as AI publications, citations, patents, hiring, skills, talent and investment. Based on these
indicators, a composite index can be generated. As regards the methodology, the index is built
as a modular weighted mean, alongside the 23 indicators – which are the so-called sub-pillars –
5
and the 3 high-level pillars on R&D, economy, and inclusion. According to the standard version
of the index, the top five countries in 2021 are the US, China, India, the UK and Canada, and
there exists a remarkable gap between the frontrunners and those lower down the scale. Fur-
thermore, the HAI annually produces an AI Index Report (Perrault et al., 2019, Zhang et al.,
2021 and 2022). It reviews AI trends in research and development, technical performance, eco-
nomic performance, education, ethical challenges, diversity, AI policies and national strategies.
The 2022 report highlights the continued rise in the number of AI publications, AI tools and
technical applications, and the continued dominance of US-China collaborations, followed some
way behind with UK-China collaborations in second position (Zhang et al., 2022). Earlier AI
Index reports indicated among others a growth in AI hiring, an increase in AI education, some
stagnating figures in terms of diversity, a rise in national AI strategies and increased government
regulation (Zhang et al., 2021, 2022).
Other indexes have also been developed by the EU and OECD. Since 2018 the European
Commission AI Watch initiative monitors i) ‘industrial, technological and research capacity as
well as policy initiatives in the Member States’ and ii) ’the uptake and technical developments
of Artificial Intelligence and its impact in the economy, society and public services’. The aim
of this data index is to ‘monitor and facilitate the implementation of the European Strategy
for AI.’ AI Watch provides information on AI developments for different countries worldwide
through an AI landscape dashboard and via further documents and analyses (Lopez Cobo et
al., 2021). It holds among others information on AI players, firms, patents and publications.
Their index covers 28 indicators that are aggregated into 6 dimensions: global view on the AI
ecosystem, industry, research and development, technology, societal aspects, and AI in the public
sector. Compared to the other measures, the AI Watch Index is planned to be constructed for
European member countries, only, but it will allow for more regional-level analyses which will
enable identifying local hubs of AI activity.
The OECD Artificial Intelligence Policy Observatory provides an ‘interactive database of AI
policies and initiatives from countries, territories and other stakeholders to facilitate interna-
tional cooperation, benchmarking and help develop best practices.’ (OECD, 2022a). Their
observatory provides real-time AI news, country dashboards, live data and visualisations on AI.
Supporting research papers analyse for example the skills required for emerging jobs in AI based
on job posting data (OECD, 2021), national policies and stakeholder initiatives (OECD 2022a).
6
Increasingly, several private consultancies have also developed similar data tools to monitor
AI developments (McKinsey 2017 and 2018). They analyse for example the degrees of AI
adoption and absorptions rates of firms and predict how these levels might develop across time.
The State of AI report (Benaich and Hogarth, 2020) reveals the development of different AI-
related components along the topics of research, talent, industry, politics, and predictions. They
show for example that China and India saw an immense growth in AI talent, and that India’s
AI research is the most diverse in terms of having the highest share of women participating
compared to other countries.
One of the newest indexes with the widest coverage of countries and indicators is the Tortoise
Global AI Index (2019, 2022). It ranks a set of 62 countries according to their scores along
seven sub-indices, based on a set of 143 variables. The countries range from industrialised to
emerging market economies and developing countries (see Table A.2 in the Appendix). This
index is the newest, most comprehensive and internationally comparative measure provided by
Tortoise Media.
The variables were selected across seven different pillars, namely according to the branches
of talent, infrastructure, operating environment, research, development, governmental strategy,
and commercial activities. The pillars were aggregated to three key areas: i. Investments, cov-
ering commercial ventures and governmental strategy; ii. Innovation, covering research and de-
velopment; and iii. Implementation, covering talent, infrastructure and operating environment.
The variables/ indicators as well as the pillars were weighted for importance by consultations
of experts, and different sub-indices were constructed according to the different pillars of the
Global AI Index, as named above.
Regarding the total overall score, the USA is ranked no. 1 and their score is normalized to 100.
In descending order follow - with a large distance to the US – China (attaining a score of 58.2),
the UK, Canada, South Korea, Israel, Germany, Netherlands, Singapore, France, Australia.
The three lowest scoring countries are Egypt, Nigeria and Pakistan. Regarding the score for the
sub-index on talent, the USA ranks first with a score of 100, followed by India with a score of
41.3, then Singapore, the UK, Israel, Netherlands, Ireland, Canada, Sweden, Norway.
Measurement of Artificial Intelligence at the micro-level is done in the past research literature
either at the level of the firm (Eurostat, 2022, Hunt et al., 2022), using job vacancy or occupa-
tional data or by generating an occupation-based AI index (Acemoglu and Restrepo 2020; Felten
7
et al., 2018, 2019), or based on patents, considering the inventions/ innovations that relate to
the elements of AI (Thielmann et al., 2021, Martinelli et al., 2021).
In terms of firm-level evidence, Hunt et al. (2022) conducted a survey on more than 750 UK
business leaders. They find evidence for both job creation and job destruction resulting from
the use of AI. Eurostat (2022) provides enterprise-level statistics, on the one hand through its
’Community survey on ICT usage and e-commerce in enterprises’. These statistics give evidence,
for example, on how many enterprises in the European countries use AI systems, or how many
enterprises analyse big data by the use of natural language processing. On the other hand,
Eurostat (2020) carried out the ’European enterprise survey on the use of technologies based on
artificial intelligence’ on a set of over 9600 enterprises, analysing AI adoption and its obstacles.
Clearly, more firm-level data - which are accessible to researchers - are needed in the future to
analyse the effects of AI for labour market outcomes (Frank et al., 2019).4
Acemoglu et al. (2020), using vacancy data from Burning Glass, find that establishments that
are AI-exposed reduce hiring in non-AI positions. However, they find no relation between AI
exposure and employment or wage growth at the occupational or industry level. Felten et al.
(2018, 2019) constructed the AI Occupational Impact index (AIOI) based on information about
occupations and abilities from O*NET and on information about tasks from Amazon Mechanical
Turk, such as image recognition, translation, or the ability to play strategic games. The authors
find for the USA evidence for employment and wage growth for high-income occupations but
none for low- or middle-income occupations. Consequently, their results indicate increasing
income inequality and labour market polarization.
Official statistics usually do not provide fixed sectoral codes or industry-related information
regarding artificial intelligence. To give an example, it is clear to identify what an automobile
is. In official statistics, one would classify this product according to a certain definition and
assign to it a sectoral code/ industry sector. For artificial intelligence, however, many different
elements are involved, as indicated in the Introduction. The various elements of AI are applied
across different industries and used for the production of different goods.
The OECD (2022b) has recently set up a framework for the classification of AI systems,
acknowledging that AI is diffusing through all sectors. Thielmann et al. (2021) classify AI
4Frank et al. (2019) describe barriers that prevent from examining the effects of AI on labour markets. Those
barriers comprise high quality data about requirements of occupations, lack of models of skill substitution or
human-machine complementarity, or a lack of modeling frameworks that bridge broader economic dynamics and
institutional processes.
8
patents based on a patent data set from the European Patent Office and by developing a new
two-step method that uses web-scraping for AI-related keywords as well as a support vector
machine and LDA topic modelling. Martinelli et al. (2021) also classified AI patents, they use
classification rules based on keywords and industry (CPC) codes.
In sum, this review indicates since around 2017 the extensive development of AI Indexes
aiming to measure and compare the growth of this technology across sectors, occupations and
countries. Here we set out to provide a deeper analysis of one of these indexes to understand
what this can tell us about new and conventional measures of its development.
3. Data
One of the interesting features of the Tortoise Global AI Index is the combination of conven-
tional and innovative new indicators sourced from new (social) media, platforms, and companies.
The variables we use for our analyses of measures of tech talent – alongside other variables for
education-based measures and other variables used in the analysis – are summarised in Table 1
and explained here. The data was made available to us for scientific analyses, it is a cross-section
and comes from the 2020 release.5
Tortoise uses information from LinkedIn to generate variables on the number of people who
describe themselves to be currently working as machine learning engineers (ML Engineers), ar-
tificial intelligence engineers (AI Engineers), Data Scientists and Data Engineers, respectively.
LinkedIn is a social network in order to gain business contacts, to conduct professional network-
ing, and to use it for job searches. Employees upload their CVs and their details about their
career paths. Employers post their job ads and the job details. More than 774 million users
from over 200 countries were registered by 2021.
Further, we use the number of AI Meetup members that Tortoise has collected. Meetup is
a social media platform for people to connect and to organise online groups for hosting virtual
or in-person events. The groups are built around similar themes or interests. The platform is
considered to be a place to meet and discuss ideas around specific topics, in this case AI.
Variables on the number of GitHub stars and GitHub commits are built from information from
GitHub. GitHub is an open-source software, hosting programming codes and data repositories.
It currently hosts more than 200 million repositories, from more than 73 million developers.
5Further information can be found in the Methodology Report (Tortoise, 2019).
9
Table 1: Description of variables
Variable Description and Measurement Data Source
Artificial Intelligence Engineers Number of Artificial Intelligence Engineers on Social Media per
million people in a country
Linkedin
Machine Learning Engineers Number of Machine Learning Engineers on Social Media per mil-
lion people in a country
Linkedin
Data Engineers Number of Data Engineers on Social Media per million people in
a country
Linkedin
Data Scientists Number of Data Scientists on Social Media per million people in
a country
Linkedin
AI Meetup members Number of Artificial Intelligence Meetup members per million peo-
ple in a country
Meetup
GitHub stars Number of GitHub stars per million people in a country GitHub
GitHub commits Number of GitHub commits per million people in a country GitHub
Kaggle Grandmasters Number of Grandmasters and Masters in Kaggle per million people
in a country
Kaggle
Enrolment in AI courses Share of enrolment in AI courses proportional to the population
in a country
Coursera
Enrolment in ML courses Share of enrolment in ML courses proportional to the population
in a country
Coursera
Researchers in STEM Number of full-time equivalent jobs as Researchers in STEM in a
given country per million people
UNESCO
IT graduates Share of IT graduates in a country proportional to all graduates UNESCO
IT undergraduates Share of IT undergraduates in a country proportional to all grad-
uates
UNESCO
Science graduates Share of Science graduates in a country proportional to all grad-
uates
UNESCO
Science undergraduates Share of Science undergraduates in a country proportional to all
graduates
UNESCO
STEM graduates Share of STEM graduates in a country proportional to all gradu-
ates
UNESCO
GDP GDP (billion) in relation per million people World Bank
Several governmental AI strategy / target
indicators
Binary indicator of whether a country has whether a country has
a national AI strategy, an AI governmental body, whether it ded-
icated money to AI, whether it has measurable AI targets, or
whether it has an AI training or upskilling strategy
National strategy
documents
Digital infrastructure 5G implementation score, number of networks per country Speedtest Ookla
5G Map
Patents Number of filed AI patents by inventors per million people Google BigQuery
R&D expenditures Proportion, total amount of public spending on research and de-
velopment in relation to total GDP
World Bank
Note: This Table shows the description, measurement and sources of variables used for the Tortoise Global AI
Index.
10
From this platform, information can be gained about the users. GitHub stars are nominated
and selected by other users. Moreover, the number of commits, that is changes to an existing
code, can be tracked. These figures can be looked up for different topics, e.g. on AI, Machine
Learning, Data Science.
Kaggle is an online community, addressing data scientists and people working in machine
learning. It is a platform that facilitates joint work of data scientists and machine learning
engineers as well as education in AI. It holds and publishes datasets and offers competitions in
data science and machine learning. Tortoise uses Kaggle to extract information on the number
of contributors to this platform that have achieved the status Grandmaster or Master.
From Coursera information is gained on the number of enrolments in courses, in particular in
AI and ML courses. Coursera is offering online-based education and courses. The courses are
not given by Coursera itself, but the platform is working together with different universities,
administering and streaming the content. There were about 92 million users of the platform in
2021.
In addition to these new types of indicators, we also use traditional education-based measures
of talent, which are the number of Researchers in STEM, the shares of IT graduates, IT un-
dergraduates, Science graduates, Science undergraduates and STEM graduates. Those figures
come from the UNESCO.
Furthermore, we use a measure of digital infrastructure, which is the 5G implementation
score per country. This indicator is the number of 5G networks per country. Tortoise gained
this variable from Speedtest. Speedtest by Ookla is an online tool that provides statistics and
analyses of regional internet access performance metrics around the world. It displays the speed
and connection coverage.
With the use of BigQuery, Tortoise generated variables from publicly available patent databases.
Google BigQuery is an analytical programming tool to manage and use big data that are avail-
able online. It is hosted on Google’s cloud platform. The way it works is that the user is making
or programming a query, e.g. searching for keywords or determining the count of words in a big
dataset.
Moreover, we use a variable of GDP and of R&D expenditures as a proportion of GDP. Those
figures come from the World Bank. Further, Tortoise generated various variables that cover
governmental AI strategies and targets, such as whether a country has a national AI strategy,
11
an AI governmental body, whether it dedicated money to AI, whether it has measurable AI
targets and whether it has an AI training or upskilling strategy. This information was gained
from national strategy documents.
4. Analysis
4.1. Comparing Talent in Europe and the World. The following Figures give an overview
of the state of AI and talent in Europe and across the World. In Figures 1 and 2, we plot the log
of the total AI Index score against the log of the talent score. Figure 1 displays the outcomes
for the European sample, Figure 2 for the World sample. What these graphs also show is the
degree of R&D expenditures in a country. Red coloured dots display half of the number of
countries that have a higher proportion of R&D expenditures per GDP than the other half of
the countries (median), blue coloured dots display a lower proportion of R&D expenditures per
GDP.
The Figures show a positive relationship between the overall Global AI Index score and the
talent score. Countries that have a higher talent score, score generally higher in the Global AI
Index. This relationship holds both for the European and the World sample. Importantly, the
graphs show that generally those countries scoring higher in talent and in the Global AI Index
tend to be the countries that have a higher proportion of R&D expenditures. Regarding Europe,
a group of forerunners in terms of overall AI and talent can be identified which comprises the UK,
Netherlands, Germany, France, Ireland, Switzerland and the Scandinavian countries Denmark,
Sweden, Finland. Those falling behind are Hungary, Slovenia, the Czech Republic, Slovakia and
Greece. From Figure 2, we can see for the World sample that the USA stands out, and mainly
countries in Africa are lagging behind, both in terms of the overall score of the Global AI Index
and in terms of the talent score.
Importantly, evaluating the scores against the 45 degree line, one can see that China actually
scores better in terms of the overall AI score than in the talent score (it lies above the 45 degree
line). The USA, however, scores somewhat better on talent than in the overall AI score (it lies
below the line). India and Norway also score comparatively well in terms of talent (lying below
the line). Many African and Asian countries, however, score comparatively lower in terms of
talent as compared to the overall AI score (lying above the line).
12
Figure 1. Global AI and Talent - Europe
Note: This Figure shows the relationship between log Global AI Index and log Talent sub-index.
The colours of the dots denote the proportion of R&Dexpenditures in GDP of the country to be
higher (red colour) or lower (blue colour) than the median across all countries. Country codes
can be found in Appendix Table A.2.
4.2. Comparing new tech talent and traditional education-based measures. Our anal-
ysis compares new tech talent measures with traditional education-based measures. Table 2 dis-
plays correlation coefficients between the different variables. A correlation is classified as ‘strong’
if the coefficient is larger than 0.75, ’moderate’ when it is between 0.5 and 0.75, and ’weak’ when
it is between 0.25 and 0.5. We can see that both for the sample of European countries and for
the whole World sample several associations can be classified as ’strong’.
This is particularly true for the relation between Researchers in STEM and AI Engineers for
the European sample, and for Researchers in STEM and ML Engineers, Data Scientists, and
Kaggle Grandmasters for the World sample.
Various relations can be classified as moderate, which are for Europe the relation between
Researchers in STEM, Data Scientists and Kaggle Grandmasters, IT graduates and ML En-
gineers, enrolment in ML courses, ML Engineers and Kaggle Grandmasters, and enrolment in
AI courses and Kaggle Grandmasters. For the World sample there are moderate relationships
for Researchers in STEM, AI Engineers, Data Engineers and GitHub commits, between STEM
13
Figure 2. Global AI and Talent - World
Note: This Figure shows the relationship between log Global AI Index and the log Talent sub-
index. The colours of the dots denote the proportion of R&Dexpenditures in GDP of the country
to be higher (red colour) or lower (blue colour) than the median across all countries. Country
codes can be found in Appendix Table A.2.
graduates and Github commits, between enrolments in ML courses, AI Engineers, ML Engi-
neers, Data Engineers, Data Scientists and Kaggle Grandmasters, and between enrolment in
AI courses, ML Engineers, Data Engineers, Data Scientists and Kaggle Grandmasters. The
correlations can also be visualised by the following Figures 3 - 5.
4.3. Detecting forerunners and laggards. With the following graphs - which we show only
for the European sample - we display correlations between variables and detect groups of coun-
tries that are forerunners or laggards in terms of talent. Countries’ new tech talent measures can
be more precisely predicted by education-based measures (we will name this ’well approximated’
in the following) when their data points lie close to a linear fit line in a scatterplot between two
variables. The farther away a country’s data point lies from the line, the less well does the new
measure approximate the old one. This relationship is mirrored by the correlation coefficients
which were discussed in the previous subsection.
We added further information to the graphs. The size of the bubbles corresponds to a country’s
GDP per capita. The colours of the bubbles denote the proportion of R&D expenditures per
14
Table 2: Correlation Matrix
AI ML Data Data AI Meetup GitHub GitHub Kaggle
Engineers Engineers Engineers Scientists members stars commits Grandmasters
European sample
Researchers 0.789 0.453 0.343 0.638 0.275 0.159 0.169 0.515
IT graduates 0.167 0.52 0.331 0.197 0.146 -0.282 -0.044 0.388
IT undergraduates 0.26 0.301 0.087 0.047 -0.009 -0.419 -0.02 0.448
Science graduates 0.039 0.11 0.074 0.137 0.482 0.235 0.048 -0.059
Science undergraduates 0.339 0.042 -0.009 0.211 0.509 0.169 0.174 0.077
STEM graduates 0.01 0.116 -0.151 -0.12 0.141 -0.002 -0.121 0.119
Enrolment in ML courses 0.233 0.53 0.371 0.282 -0.016 -0.421 -0.222 0.654
Enrolment in AI courses 0.085 0.446 0.382 0.203 -0.013 -0.452 -0.207 0.666
World sample
Researchers 0.704 0.794 0.674 0.766 0.444 0.038 0.725 0.789
IT graduates 0.221 0.319 0.263 0.269 0.091 -0.017 0.27 0.32
IT undergraduates 0.325 0.375 0.269 0.269 0.15 -0.143 0.355 0.409
Science graduates 0.374 0.367 0.214 0.3 0.201 0.227 0.392 0.093
Science undergraduates 0.408 0.307 0.167 0.263 0.173 0.24 0.373 0.121
STEM graduates 0.39 0.441 0.286 0.373 0.16 0.184 0.515 0.36
Enrolment in ML courses 0.572 0.679 0.635 0.646 0.47 -0.28 0.387 0.641
Enrolment in AI courses 0.545 0.638 0.639 0.637 0.476 -0.35 0.349 0.665
Note: This Table shows correlation coefficients between the variables. The variables are logged. For an explanation
of their measurement see Table 1. The European sample contains the following countries: Austria, Belgium, Czech
Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy, Lithuania,
Luxembourg, Malta, Netherlands, Norway, Poland, Portugal, Slovakia, Slovenia, Spain, Sweden, Switzerland,
United Kingdom. The full list of countries (World sample) can be found in Appendix Table A.2.
GDP of the country to be higher (red colour) or lower (blue colour) than the median across all
countries.
Figure 3 shows the relationship between Researchers in STEM and AI Engineers. All measures
are logged and given per million people in a country. What we see from Figure 3 is a strong
correlation between Researchers in STEM and AI Engineers. Forerunners, both in terms of the
number of Researchers in STEM and in terms of new tech talent, are the Scandinavian countries
- Sweden, Denmark, Norway, Finland, Iceland - and the Netherlands. Laggards are Hungary,
Spain, Greece, Poland, Lithuania, Slovakia and in particular Italy and Malta. What the graph
15
Figure 3. AI engineers and Researchers in STEM
Note: This Figure shows the relationship between log Researchers in STEM per capita and log
AI engineers per capita. The size of the bubbles corresponds to a country’s GDP per capita.
The colours of the bubbles denote the proportion of R&Dexpenditures in GDP of the country
to be higher (red colour) or lower (blue colour) than the median across all countries. Country
codes can be found in Appendix Table A.2.
also reveals is that the higher is GDP per capita and the proportion of R&D expenditures per
GDP of a country, the higher is the number of researchers and new tech talent workers.
Figure 4 shows the relationship between Data Scientists and Researchers in STEM. The
relation is positive and we can see that the higher are the proportion of R&D expenditures
per GDP and GDP per capita the higher is the number of Researchers in STEM and of Data
Scientists. Forerunners in terms of Data Scientists are the Scandinavian countries, Iceland,
Luxembourg, Netherlands, Switzerland, Ireland. Laggards are Italy, Malta, Slovakia, Poland,
Hungary and Greece.
Figure 5 shows the relationship between IT graduates and ML Engineers. We can see that the
correlation is weaker than it was for the other two relationships explained before, the data points
lie farther away from the linear fit line. Moreover, there is no clear relationship towards the
proportion of R&D expenditures per GDP and the size of GDP per capita anymore. Forerunners
are Ireland, Estonia, Denmark and Finland. Lagging behind are Italy and Portugal. A high
16
Figure 4. Data scientists and Researchers in STEM
Note: This Figure shows the relationship between log Researchers in STEM per capita and log
Data scientists per capita. The size of the bubbles corresponds to a country’s GDP per capita.
The colours of the bubbles denote the proportion of R&Dexpenditures in GDP of the country
to be higher (red colour) or lower (blue colour) than the median across all countries. Country
codes can be found in Appendix Table A.2.
value regarding ML Engineers have Sweden, Switzerland, Ireland, Estonia, Denmark, Finland
and the Netherlands. A low value regarding ML Engineers have Italy, Portugal, Iceland, Austria
and Spain. Regarding IT graduates, Ireland, Denmark, Estonia, Finland, Poland and Malta rank
high. Regarding IT graduates, Italy, Portugal and Luxembourg and Belgium rank low.
4.4. Governments’ AI strategies and targets. Aside from the relevance of countries’ R&D
expenditures, we are interested in analysing how governmental AI strategies and targets are
related to the number of new tech talent workers. The Tortoise dataset holds several binary
variables, namely whether the government has a national AI strategy, whether it has an AI
governmental body, whether it publicly dedicated money to AI, whether it has measurable AI
targets, and whether it has a training or upskilling strategy related to AI. For the following
analyses we considered all countries given by the World sample.
Based on these binary variables, we can separate the countries into two groups: those having
an AI strategy/ target (1) or not (0). We can then test whether, for example, the volume of
17
Figure 5. ML engineers and IT graduates
Note: This Figure shows the relationship between log IT graduates and log ML engineers per
capita. The size of the bubbles corresponds to a country’s GDP per capita. The colours of the
bubbles denote the proportion of R&Dexpenditures in GDP of the country to be higher (red
colour) or lower (blue colour) than the median across all countries. Country codes can be found
in Appendix Table A.2.
tech talent differs across the groups of countries that have a national AI strategy versus those
that have no national AI strategy.
Instead of just focussing on one moment of the distribution of tech talent workers – which
is what a simple t-test would do, focussing on the mean value across the two groups - we take
the Kolmogorov-Smirnov test which considers all moments of the distribution and which tests
for stochastic dominance of the distribution of tech talent workers for one group (e.g. the
group of countries that have a national AI strategy) over the distribution of tech talent workers
for the other group (the group of countries that have no national AI strategy). Formally, let
F1 denote the cumulative distribution function of tech talent workers for one group, and F2
the cumulative distribution function for tech talent workers for the other group. First order
stochastic dominance of F1 relative to F2 then means that F1(x) - F2(x) must be less or equal
to zero for all values of x, with strict inequality for some x.
18
Table 3: New tech talent and AI targets - Non-parametric distribution tests
H1 H2 H3
Artificial Intelligence Engineers
1. Government has dedicated National AI Strategy versus has none 0.185 0.092 0.918
2. Government has dedicated AI Governmental Body versus has none 0.054 0.027 0.954
3. Government has publicly dedicated Money to AI versus has none 0.103 0.052 0.962
4. Government has measurable AI targets versus has none 0.185 0.092 0.918
5. Government has AI Training/Upskilling Strategy versus has none 0.059 0.029 0.929
Machine Learning Engineers
1. Government has dedicated National AI Strategy versus has none 0.042 0.021 0.857
2. Government has dedicated AI Governmental Body versus has none 0.031 0.015 0.819
3. Government has publicly dedicated Money to AI versus has none 0.013 0.006 0.863
4. Government has measurable AI targets versus has none 0.042 0.021 0.857
5. Government has AI Training/Upskilling Strategy versus has none 0.074 0.037 0.809
Data Scientists
1. Government has dedicated National AI Strategy versus has none 0.025 0.012 0.980
2. Government has dedicated AI Governmental Body versus has none 0.041 0.020 0.861
3. Government has publicly dedicated Money to AI versus has none 0.018 0.009 0.945
4. Government has measurable AI targets versus has none 0.025 0.012 0.980
5. Government has AI Training/Upskilling Strategy versus has none 0.071 0.035 0.963
Data Engineers
1. Government has dedicated National AI Strategy versus has none 0.025 0.012 1.000
2. Government has dedicated AI Governmental Body versus has none 0.332 0.167 0.898
3. Government has publicly dedicated Money to AI versus has none 0.020 0.010 0.999
4. Government has measurable AI targets versus has none 0.025 0.012 1.000
5. Government has AI Training/Upskilling Strategy versus has none 0.039 0.019 1.000
Github stars
1. Government has dedicated National AI Strategy versus has none 0.686 0.360 0.593
2. Government has dedicated AI Governmental Body versus has none 0.567 0.291 0.955
3. Government has publicly dedicated Money to AI versus has none 0.199 0.100 0.645
4. Government has measurable AI targets versus has none 0.686 0.360 0.593
5. Government has AI Training/Upskilling Strategy versus has none 0.556 0.285 0.752
Note: This Table displays results from a non-parametric Kolmogorov-Smirnov test for the equality of distribution
across governmental AI strategies. The tested hypotheses are: H0: the distribution for the two groups are
identical. H1: the distributions of the two groups differ. H2: The distribution of the first group (0) is first-order
stochastically dominated by the distribution of the second group (1). H3: The distribution of the second group is
first-order stochastically dominated by the distribution of the first group. A p-value of 0.05 or smaller indicates
that the null-hypothesis can be rejected in favour of the alternative hypothesis H1, H2 or H3 at an error level of
5 percent or better.
19
Table 3 shows the results. In the case of Machine Learning Engineers and Data Scientists we
can see that the null hypothesis that the distributions of tech talent workers are equal for the
two groups of countries can be rejected in favour of the alternative hypothesis (H1) that the two
distributions differ and in favour of the alternative hypothesis (H2) that the distribution for the
first group is first order stochastically dominated by the distribution of the second group, given
that the p-value is smaller than 0.1. This means that those countries that have a national AI
strategy, that have an AI governmental body, publicly dedicated money to AI, have measurable
AI targets, and have an AI training/ upskilling strategy, have a higher number of Machine
Learning Engineers and Data Scientists.
The same generally also holds for the number of AI Engineers and Data Engineers. However,
for the case of Data Engineers, no significant difference in the distributions for those countries
that have an AI governmental body versus those that have none can be found. Moreover, for
the case of AI Engineers no significant differences at conventional levels can be found between
countries that have a national AI strategy, publicly dedicated money to AI or have measurable
AI targets, as the p-values lie above 0.1. However, the hypothesis H0 is rejected in favour of the
hypothesis H2 of first order stochastic dominance. Aside from these relationships, no significant
relationship towards the difference of the distributions of countries can be found for the number
of Github stars.
4.5. Talent, Digital Infrastructure, and Innovation. We further analysed the relation
between talent and innovation activities. Our analyses are restricted by the small amount of
observations, however. Thus, our analyses can only be seen as some tentative evidence for the
macroeconomic relationships across countries. We consider the World sample for our analyses.
The results are shown in Table 4.
We estimate the following equation:
Innovationi=β0+β1Talenti+β2Xi+δi+i
where innovation is measured by the number of filed AI patents by inventor per million capita,
talent is measured each by different traditional education-based measures (Researchers in STEM,
STEM graduates, IT undergraduates, IT graduates, Science undergraduates, Science graduates)
20
as well as new tech talent measures (Data Scientists, Data Engineers, ML Engineers, AI Engi-
neers, Github stars, Github commits, AI Meetup members), X is a vector of additional explana-
tory variables, namely GDP per million capita, and digital infrastructure which is measured as
the 5G implementation score, δis a set of continental fixed effects to capture further variation
across countries and is the error term.
The results show that a positive and significant link can be found between Science under-
graduates, Science graduates, Github stars and innovative activities. Importantly, digital in-
frastructure is positively and significantly related to innovative activities. The effect is found in
every equation, which provides robust evidence. According to the results, a 1 percent increase
in the score of implementation of 5G is associated with an increase in the number of AI patent
applications by an inventor by about 0.6 percent on average.
5. Conclusion
In this contribution, we reviewed how AI development is measured and assessed in the liter-
ature. Our aim was to detect what we can learn from the recent attempts to measure AI. A
special focus has been given to the talent component of AI. We utilised the richness of variables
available with a new dataset based on the Tortoise Global AI Index and compared new measures
of tech talent to traditional education-based measures.
Our analyses show that several new tech talent measures are strongly correlated with tradi-
tional measures. In particular, measures of AI Engineers and of Data Scientists are displaying
a strong connection to measures relating to the number of Researchers in STEM and to the
number of IT graduates. With regard to Europe, we further detected countries that belong to
the forerunners (Western European and Scandinavian countries) or laggards (South and Eastern
European countries) in terms of global AI, tech talent, and skill-based measures. Our analyses
reveal a relationship between governments’ R&D expenditures, the number of Researchers in
STEM and new tech talent measures. A further relation exists between countries’ AI strategies
and targets and the number of new tech talent workers. Moreover, we reveal a link between tal-
ent, digital infrastructure in terms of 5G implementation, and innovative activities as measured
by filed AI patents by inventors.
Based on these results, we find that attempts to measure AI based on composite indices do
a good job in terms of providing new evidence and to allow for comparisons across countries.
21
Table 4: Talent, Digital Infrastructure, and Innovation - Regression analysis
Dep var: Number of AI patents (1) (2) (3) (4) (5) (6) (7)
Traditional education-based measures
Digital infrastructure 0.5886** 0.6545** 0.6175** 0.6796** 0.6625** 0.4299* 0.5212**
GDP -0.162 -0.0795 -0.454 -0.5279 -0.4912 -0.1884 -0.2833
Researchers in STEM -0.7566
STEM graduates 0.155
IT undergraduates -0.5192
IT graduates -0.4319
Science undergraduates 1.4206**
Science graduates 0.8632*
Continental fixed effects yes yes yes yes yes yes yes
R20.306 0.341 0.374 0.381 0.38 0.429 0.401
n 48 40 43 43 43 43 43
Dep var: Number of AI patents (8) (9) (10) (11) (12) (13) (14)
New tech talent measures
Digital infrastructure 0.5836** 0.5993** 0.5728** 0.5853** 0.4382* 0.5643** 0.717**
GDP -0.1148 -0.257 -0.0164 -0.2637 -0.1317 -0.3102 -0.5395
Data Scientists -0.0528
Data Engineers 0.1231
ML Engineers -0.2226
AI Engineers 0.202
Github stars 0.4640**
Github commits 0.2758
AI Meetup members 0.0262
Continental fixed effects yes yes yes yes yes yes yes
R20.306 0.307 0.31 0.299 0.397 0.329 0.358
n 48 48 48 47 47 48 42
Note: This Table displays results from OLS regressions. The dependent variable is the number of filed AI patents
by inventors. Digital infrastructure is measured as the level of 5G implementation in a country. All variables
are logged. Robust standard errors were computed. ** denotes significance at the 5 percent level, * denotes
significance at the 10 percent level.
Based on newly available data sources through (social) media or other platforms, informative
new tech talent measures can be generated. Those new tech talent measures prove to be useful,
as they give real-time information on the current state of talent of a country, as well as for the
development potential of jobs and labour markets in the future. On the other hand, a result of
our analyses is that measures of researchers and students in STEM or IT predict the numbers of
new tech talent well. In other words, if politicians want to address new tech talent in AI, Data
22
Science or Machine Learning, then it is important to view figures of undergraduates, graduates
and researchers in STEM, science or IT. Educational provision of STEM or IT qualifications
are crucial to the growth of tech talent, and some countries are striding ahead while others are
lagging behind with crucial consequences for different patterns of economic growth.
Our study might be limited by the reliability of the information used to construct these
innovative variables. For example, one could ask whether employees will enter true information
into their LinkedIn profiles. Moreover, provision of information in those platforms might be
restricted by a country’s law and regulations (e.g. China). It might be less of an issue for our
analyses, though, as the variables we use relate to very professional information, only, and no
information regarding religious or political views is required.
In terms of policy implications, our results hint at the relevance of providing R&D expenditures
and of fostering AI strategies and targets to support new jobs for researchers and to stimulate
the emergence of new tech talent in a country. Moreover, we can reveal tentative evidence that
digital infrastructure and innovative activities are linked to each other, indicating that a good
quality of the digital infrastructure is needed in order to support the emergence of innovative
activities.
23
References
Acemoglu, D.; Autor, D.; Hazell, J.; Restrepo, P. (2020). AI and jobs: Evidence from Online
Vacancies, NBER working paper series, No. 28257.
Acemoglu, D. and Restrepo, P. (2019). The wrong kind of AI? Artificial intelligence and the
future of labour demand. Cambridge Journal of Regions, Economy and Society. Available at:
https://pdfs.semanticscholar.org/6870/9c00115b93e846e1168e751a6cfc9cef7ac7.pdf.
Acemoglu, D.; Restrepo. P. (2018). Artificial Intelligence, Automation, and Work, NBER
working paper series, No. 24196.
Adams-Prassl, A. (2020). The Gender Wage Gap in an Online Labour Market: The Cost of
Interruptions, CEPR Discussion Paper No. DP14294.
Agrawal, A., Gans, J.S. and Goldfarb, A. (2019). Artificial Intelligence: The Ambiguous Labor
Market Impact of Automating Prediction. Journal of Economic Perspectives, 33(2), pp.31–50.
Agrawal, A.; Gans, J.; Goldfarb, A. (2016). The Simple Economics of Machine Intelligence,
Harvard Business Review, Available at: https://hbr.org/2016/11/the-simple-economics-of-
machine-intelligence.
Benaich, N. and Hogarth, I. (2020). “State of AI Report 2020”, Available at: https://www.
stateof.ai/2020.
Black SE; Muller C; Spitz-Oener, A; He Z; Hung, K; Warren JR (2021). The importance of
STEM: High-school knowledge, skills and occupations in an era of growing inequality, Research
Policy, 50(7), 104249.
Boden, M. A. (2016). AI - Its nature and future, Oxford University Press.
Brynjolfsson, E.; Mcafee, A. (2014). The Second Machine Age: Work, Progress, and Prosperity
in a Time of Brilliant Technologies. New York: W.W. Norton Company.
Cedefop (2018). Insights into skill shortages and skill mismatch: Learning from Cedefop’s
European skills and jobs survey, 3075, Available at: https://www.cedefop.europa.eu/en/
publications/3075#group-details.
Coursera, https://www.coursera.org.
European Commission (2021a). European Skills Agenda: progress on the 12 flagship actions,
Available at: https://ec.europa.eu/social/main.jsp?catld=1223langld=en
European Commission (2021b). Coordinated Plan on Artificial Intelligence 2021 Review, Avail-
able at: https://digital-strategy.ec.europa.eu/en/library/coordinated-plan-artificial-intelligence-
2021-review?msclkid=af85c736b07d11ecb8544e9d2b55e559
European Commission (2020). White paper on Artificial Intelligence: a European approach to
excellence and trust, ,Available at: https://ec.europa.eu/info/sites/default/files/commission-
white-paper-artificial-intelligence-feb2020en.pdf
24
European Commission (2018). Communication from the Commission to the European Parlia-
ment, the European Council, the Council, the European Economic and Social Committee
and the Committee of the Regions - Artificial Intelligence for Europe, COM/2018/237 final,
Available at: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=COM:2018:237:FIN.
Eurostat (2022). Community survey on ICT usage and e-commerce in enterprises, Available at:
https://ec.europa.eu/eurostat/web/products-eurostat-news/-/ddn-20210413-1
Eurostat (2020). European enterprise survey on the use of technologies based on artificial
intelligence, https://digital-strategy.ec.europa.eu/en/library/european-enterprise-survey-use-
technologies-based-artificial-intelligence
Felten, E; Raj, M; Seamans, R. (2019). The Occupational Impact of Artificial Intelligence: La-
bor, Skills, and Polarization, Available at: https://papers.ssrn.com/sol3/papers.cfm?abstractid=
3368605.
Felten EW, Raj M, Seamans R (2018) A Method to Link Advances in Artificial Intelligence to
Occupational Abilities. AEA Papers and Proceedings 108, pp.54–57.
Frank, M., Autor, D., Bessen, J., Brynjolfsson, E., Cebrian, M., Deming, D., Feldman, M.,
Groh, M., Lobo, J., Moro, E., Wang, D., Youn, H., Rahwan, I., 2019. Toward understanding
the impact of artificial intelligence on labor. PNAS, 116 (14), pp. 6531–6539.
Frey, C.B. (2019). The Technology Trap - Capital, Labor, and Power in the Age of Automation,
Princeton University Press, Princeton, New Jersey.
Frey, C.B. and Osborne, M.A. (2017). The future of employment: How susceptible are jobs to
computerisation? Technological Forecasting and Social Change, 114(1), pp.254–280.
GitHub, https://github.com.
Google BigQuery, https://cloud.google.com/bigquery.
Graetz, G. and Michaels, G. (2018). Robots at Work. The Review of Economics and Statistics,
100(5), pp.753–768.
Harrigan, J., Reshef, A. and Toubal, F. (2020). The March of the Techies: Job Polarization
Within and Between Firms. Research Policy, p.104008.
Hunt, W., Sarkar, S. and Warhurst, C. (2022). Measuring the impact of AI on jobs at the
organization level: Lessons from a survey of UK business leaders. Research Policy, 51(2),
p.104425.
Kaggle, https://www.kaggle.com.
Karabarbounis, L. and Neiman, B. (2013). The Global Decline of the Labor Share. The Quar-
terly Journal of Economics, 129(1), pp.61–103.
K¨assi, O. and Lehdonvirta, V. (2018). Online labour index: Measuring the online gig economy
for policy and research. Technological Forecasting and Social Change, 137, pp.241–248.
25
Korinek, A., Stiglitz, J. (2017). Artificial intelligence and its implications for income distribution
and unemployment, NBER working paper series, No. 24174.
Krenz, A.; Prettner, K.; Strulik, H. (2021). Robots, Reshoring, and the Lot of Low-Skilled
Workers, European Economic Review, 136, 103744.
Krenz, A.; Strulik, H. (2022). Automation and the Fall and Rise of the Servant Economy,
Available at https://papers.ssrn.com/sol3/papers.cfm?abstractid= 4071751
LinkedIn, https://www.linkedin.com.
Lopez Cobo, M., Righi, R., Samoili, S., Vazquez-Prada Baillet, M., Cardona, M. and De Prato,
G. (2021). AI Watch Index: Policy relevant dimensions to assess Europe’s performance in ar-
tificial intelligence, EUR 30681 EN, Publications Office of the European Union, Luxembourg,
2021, ISBN 978-92-76-36359-0, doi:10.2760/506611, JRC124424.
Martinelli, A., Mina, A., Moggi, M. (2021). The enabling technologies of industry 4.0: Exam-
ining the seeds of the fourth industrial revolution, Industrial and Corporate Change, 30(1),
pp.161–188.
McKinsey Global Institute (2017). Artificial Intelligence - The next Digital Frontier?, Available
at: https://www.mckinsey.com/~/media/McKinsey/Industries/Advanced\%20Electronics/
Our\%20Insights/How\%20artificial\%20intelligence\%20can\%20deliver\%20real
\%20value\%20to\%20companies/MGI-Artificial-Intelligence-Discussion-paper.ashx
McKinsey Global Institute (2018). Modeling the impact of AI on the world economy, Available
at: https://www.mckinsey.com/~/media/McKinsey/Featured\%20Insights/Artificial\
%20Intelligence/Notes\%20from\%20the\%20frontier\%20Modeling\%20the\%20impact\
%20of\%20AI\%20on\%20the\%20world\%20economy/MGI-Notes-from-the-AI-frontier-
Modeling-the-impact-of-AI-on-the-world-economy-September-2018.ashx
Meetup, https://www.meetup.com.
Mueller, H., Ouimet, P., Simintzi, E., (2017). Wage inequality and firm growth, American
Economic Review, 107 (5), pp.379–383.
Neufeind, M.; O’Reilly, J.; Ranft, F. (2018). Work in the Digital Age, London: Rowman and
Littlefield International Ltd.
OECD (2022b). OECD Framework for the Classification of AI systems, Available at: https://
www.oecd.org/publications/oecd-framework-for-the-classification-of-ai-systems
-cb6d9eca-en.htm
OECD (2022a). OECD AI Policy Observatory, https://oecd.ai/en.
OECD (2021). The human capital behind AI: jobs and skills demand from online job postings,
OECD Science, Technology and Innovation Policy Papers, No. 120.
26
Perrault, R., Shoham, Y.Y., Brynjolfsson, E., Clark, J., Etchemendy, J., Grosz, B., Lyons, T.,
Manyika, J., Mishra, S., Niebles, J.C. (2019). The AI Index 2019 Annual Report. AI Index
Steering Committee, Human-Centered AI Institute, Stanford University, Stanford, CA.
Raj, M. and Seamans, R. (2018). AI, labor, productivity, and the need for firm-level data,
NBER working paper series, No. 24239.
Russell, S.J.; Norvig, P. (2022). Artificial Intelligence - A Modern Approach, Pearson Education
Limited, Harlow, United Kingdom.
Shoham, Y., Perrault, R., Brynjolfsson, E., Clark, J., Manyika, J., Niebles, JC., Lyons, T.,
Etchemendy, J., Grosz, B. and Bauer, Z. (2018) ”The AI Index 2018 Annual Report”, AI
Index Steering Committee, Human-Centered AI Initiative, Stanford University, Stanford,
CA.
Susskind, D. (2020). A World Without Work: Technology, Automation and How We Should
Respond, Picador, Metropolitan Books, Henry Holt and Company, New York.
Speedtest Ookla, https://www.speedtest.net.
Thielmann, A.; Weisser, C.; Krenz, A. (2021). One-Class Support Vector Machine and LDA
Topic Model Integration - Evidence for AI Patents, Studies in Computational Intelligence, 981,
Soft Computing: Biomedical and Related Applications, Editors N.H. Phuong, V. Kreinovich,
Springer.
Tortoise Media (2019). The Global AI Index - Methodology, Available at: https://www.
tortoisemedia.com/wp-content/uploads/sites/3/2019/12/Global-Ai-Index-\%E2\%80\
%93-Methodology.pdf.
Tortoise Media (2022), https://www.tortoisemedia.com/intelligence/global-ai.
Vannuccini, S., Prytkova, E. (2020). Artificial intelligence’s new clothes? From general pur-
pose technology to large technical system, in Kanza, S., Frey, J.G., Niranjan, M., Hooper,
V. (Eds.), AI3SD Summer Seminar Series 2020, Southampton, UK, Available at: https://
doi.org/10.5258/SOTON/P0051.
Wooldridge, M. (2018). Artificial Intelligence, Ladybird Books Ltd, London.
Zhang, D., Mishra, S., Brynjolfsson E., Etchemendy, J., Ganguli, D., Grosz, B., Lyons, T.,
Manyika, J., Niebles, J.C., Sellitto, M., Shoham, Y., Clark, J., and Perrault, R., (2021) “The
AI Index 2021 Annual Report,” AI Index Steering Committee, Human-Centered AI Institute,
Stanford University, Stanford, CA.
Zhang, D., Maslej, N., Brynjolfsson, E., Etchemendy, J., Terah Lyons, Manyika, J., Ngo, H.,
Niebles, JC, Sellitto, M., Sakhaee, E., Shoham, Y., Clark, J., and Perrault,R. (2022). “The
AI Index 2022 Annual Report,” AI Index Steering Committee, Stanford Institute for Human-
Centered AI, Stanford University.
27
Appendix
Table A.1: Descriptive Statistics
Variable Mean Std.Dev. Min Max Obs
European sample
Researchers 4629.015 1545.99 1710 7971.821 26
IT undergraduates .1981811 .1037045 .0515769 .4914644 23
Science undergraduates .2592736 .1106549 .0699233 .5401778 23
IT graduates .0373819 .0195062 .006173 .0937091 23
Science graduates .0602543 .0350411 .0166635 .1675476 23
STEM graduates .223054 .0692773 .0488978 .3373526 23
AI Engineers 9.4516 9.559039 0 37.70651 26
ML Engineers 15.7493 8.08641 4.179104 38.77551 26
Data Engineers 49.61178 29.5439 12.29358 111.2903 26
Data Scientists 100.2944 58.76045 26.6055 242.3543 26
AI Meetup members 456.0496 596.5768 0 2380.972 26
GitHub stars 10.77434 19.66292 0 79.29037 26
GitHub commits 1042.143 1118.555 47.10183 5050.923 26
Kaggle Grandmasters .1111142 .1253225 0 .450197 26
Enrolments in ML courses .0083081 .0070471 .0009735 .0270289 20
Enrolments in AI courses .0357655 .06369 .0030001 .2982187 20
GDP 44088.11 25148.22 15595.4 114740.6 26
Government has national AI strategy 0.7307692 0.4523443 0 1 26
Digital infrastructure 274.1538 614.7071 0 2750 26
Patents 83.96525 324.9069 0 1658.33 26
World sample
Researchers 3197.113 2242.852 85.51847 7971.821 52
IT undergraduates .195248 .0969109 .0449384 .4914644 54
Science undergraduates .2574982 .1186589 .0294448 .5401778 54
IT graduates .0393755 .0220242 .006173 .1062615 54
Science graduates .0567142 .0319932 .0053736 .1675476 54
STEM graduates .2097335 .0909958 .0478209 .4400876 54
AI Engineers 5.716304 7.697737 0 37.70651 62
ML Engineers 12.55228 19.39176 .4027639 143.6464 62
Data Engineers 34.34422 33.56061 1.597333 137.8947 62
Data Scientists 63.90301 67.05617 .8585411 280.7018 62
AI Meetup members 460.1003 930.2607 0 5949.217 62
GitHub stars 104.3961 676.0236 0 5330.46 62
GitHub commits 556.2049 875.1012 0 5050.923 62
Kaggle Grandmasters .0804203 .1177718 0 .5320092 62
Enrolment in ML courses .0056305 .0065978 .0001417 .0270289 48
Enrolment in AI courses .0210933 .043893 .0003216 .2982187 48
GDP 30850.18 25252.84 1284.7 114740.6 62
Government has national AI strategy .6129032 .4910624 0 1 62
Digital infrastructure 258.5323 1076.438 0 7958 62
Patents 73.07791 287.4222 0 1658.33 62
Note: This Table displays descriptive statistics for the European sample (see Appendix Table A.2) and for the
World sample.
28
Table A.2: List of countries
Country Code European sample World sample
Argentina ARG √
Armenia ARM √
Australia AUS √
Austria AUT √ √
Bahrein BHR √
Belgium BEL √ √
Brazil BRA √
Canada CAN √
Chile CHL √
China CHN √
Colombia COL √
Czech Republic CZE √ √
Denmark DNK √ √
Egypt EGY √
Estonia EST √ √
Finland FIN √ √
France FRA √ √
Germany DEU √ √
Greece GRC √ √
Hong Kong HKG √
Hungary HUN √ √
Iceland ISL √ √
India IND √
Indonesia IDN √
Ireland IRL √ √
Israel ISR √
Italy ITA √ √
Japan JPN √
Kenya KEN √
Lithuania LTU √ √
Luxembourg LUX √ √
Malaysia MYS √
Malta MLT √ √
Mexico MEX √
Morocco MAR √
Netherlands NLD √ √
New Zealand NZL √
Nigeria NGA √
Norway NOR √ √
Pakistan PAK √
Poland POL √ √
Portugal PRT √ √
Qatar QAT √
Russia RUS √
Saudi Arabia SAU √
Singapore SGP √
Slovakia SVK √ √
Slovenia SVN √ √
South Africa ZAF √
South Korea KOR √
Spain ESP √ √
Sri Lanka LKA √
Sweden SWE √ √
Switzerland CHE √ √
Taiwan TWN √
Tunisia TUN √
Turkey TUR √
United Arab Emirates ARE √
United Kingdom GBR √ √
USA USA √
Uruguay URY √
Vietnam VNM √
Note: This Table displays the country samples and country codes.
29
