International mobility of AI scientists

Artificial Intelligence (AI) is on the rise. It is currently regarded as the key technology that will radically change our society and economy. People often speak of it as being an AI race: developments are moving fast and countries are investing heavily. Countries that are currently unable to keep up with international developments in the field of AI would put their economies at a disadvantage (see for example AINED, 2018; Ministry of Economic Affairs, Agriculture and Innovation, 2019; NRC, 2019; House of Representatives, 2020). China and the United States, in particular, are mentioned as countries that are far ahead of the troops. Many concerns are expressed about the position of the Netherlands: that investments are lagging behind neighbouring countries; that the Netherlands is training too little AI talent and that the talent is leaving en masse abroad (AINED, 2018; Kickstart AI, 2019; Ministry of Education, Culture and Science, 2019; House of Representatives, 2020).

But what is the actual situation in the Netherlands? In a series of publications, the Rathenau Instituut provides numerical insight into a number of conditions for the development of AI in the Netherlands. We look in particular at AI research. In this first fact sheet we discuss the mobility of AI scientists. We compare the Netherlands with the ten countries that had the highest publication output in the field of AI in the period 2009-2018. 

Over the period 2009-2018, the Netherlands ranks twentieth in terms of the number of AI publications. Researchers affiliated with Dutch institutions were involved in 1.2% of the worldwide AI publications from this period. China and the US lead this list and are involved in 25.6% and 16.5%, respectively, of global AI publication output in this period. This is followed by India with 8.3% and the UK with 5.4% (Elsevier AI resource centre, data generated for Rathenau Instituut). 

Looking at the period 2013-2018, AI publications represent 1.3% of the total Dutch publication output. This is lower than the share of AI publications in the total output of the EU-27: 1.8%. However, the number of AI publications in the Netherlands increased sharply in this period, by 115%. Of the ten countries with which we compare the Netherlands in this analysis, only Japan and the US experienced a larger increase. 

The citation impact of publications involving scientists from Dutch institutions is very high at 2.08 on average. This means that these AI publications are quoted more than twice as often as the global average in their field. Of the ten countries where most AI publications originate, only the US, Canada and the UK have a higher average citation score for AI publications. Compared to other countries, Dutch AI research is therefore of high quality and relevance.

For all countries, the citation impact of AI publications is higher than the average citation impact of all publications involving researchers from that country. This mainly reflects the high relevance of the subject of AI at the moment (Aksnes, Langfeldt and Wouters, 2019). This is also reflected in the development of the citation impact score. For the entire period 2009-2018, the average citation impact of AI publications in the Netherlands was 1.55 (Elsevier AI resource centre). For the period 2013-2018 it rose to 2.08. We also see a similar increase in other countries in this analysis.

As the figure below shows, the Netherlands has a high circulation of AI scientists. The figure shows the composition of the group of AI scientists who have been associated with a Dutch institution for a shorter or longer period of time over the past twenty years (1996-2019). It also looks at the composition of this group in the ten countries with the most AI publications and in the EU-27. 

Of the AI scientists connected to the Netherlands in the period 1996-2019, only 26% were not internationally mobile. 35% are AI scientists who spent a short period of time in the Netherlands (less than two years). The proportion of inflow and outflow of scientists (for a period of more than two years) is more or less the same, with 13% and 14% respectively. A similar composition of mobility categories can be seen in the United Kingdom, Canada, and France. 

AI scientists associated with the Netherlands have an average of twelve years of experience, measured by the number of years they publish articles (in Scopus). This makes them just as experienced as the AI scientists associated with the US. China, India and Iran have a relatively less experienced AI research population (nine to ten years). The AI scientists associated with France, Spain, and Japan have an average of fourteen years of experience. 

If we compare the composition of the group of AI scientists associated with the Netherlands with that of the ten most publishing countries, it is striking that the Netherlands has quite a few AI scientists who stay in the Netherlands for a short period of time (less than two years). The proportion of inflow scientists staying longer than two years is also high compared to other countries. Apart from the Netherlands, this only exceeds 10% in Canada and the United Kingdom (for the US it is 10%).

The share of departing scientists is also relatively high for the Netherlands. Only Canada, France, and the UK have a percentage that is even higher, above 15%. For Iran it's similar to the Netherlands. For the Netherlands, the proportion of departing scientists is almost equal to the number of incoming scientists. Apart from the Netherlands, this only applies to China and India. For some countries, the share of departing scientists is more than five percentage points higher than the share of incoming researchers (Canada, Germany, France, Iran, Japan and Spain). For the US alone, the share of incoming scientists exceeds the share of departing scientists.

The image that the Netherlands is a country with a lot of exchange of AI scientists is confirmed when we look at the composition of the long-term inflow and outflow of AI scientists in the figures below. For each country, the proportion of inflow and outflow of scientists is broken down into 'new' migrants and 'returnees'. New' migrants are those who, viewed from the country in question, are leaving or entering the country for the first time. 'Inflow returnees', on the other hand, make a circle. Inflow returnees are those who have left the country for the first time and return later. Outflow returnees make the opposite movement: they have first come to the country in question, and leave again more than two years later.

The first graph shows that the group of departing scientists for the Netherlands is largely made up of AI scientists who have been mobile before. Almost half of the departing scientists (47%) are AI scientists who previously came to the Netherlands from abroad. Besides the Netherlands, this only applies to the US (56%) and the UK (46%). 

If we look at the incoming scientists (second graph) we see just the opposite picture. Only 23% of them were previously settled in the Netherlands. This is higher for most countries: from 33% in Germany to 54% in India. Only in the US, the UK and Canada is this proportion lower. This difference is mainly due to the high proportion of new arrivals in the Netherlands. The share of returnees in the total of AI scientists connected to a country is between 1.5 and 3% for almost all countries. 

Finally, the third graph sets the inflow against the outflow. This shows the balance between these two flows, which we discussed above. 

The average citation impact score of AI scientists who have spent a period in the Netherlands over the past twenty years is high: 2.21. This puts the Netherlands on an equal footing with Germany. Higher citation impact scores are only achieved by AI scientists in the US (2.38), Canada (2.26), and the UK (2.24). Because this is an indication of the quality of scientists, all their publications are included, including non-AI-related publications. As a result, these citation impact scores differ from those in the first figure in this fact sheet. 

Figure 4 shows the average citation impact score of the AI scientists per mobility category (y-axis) compared to the relative productivity of those scientists (x-axis). The relative productivity is an index, based on the number of scientific articles produced annually by an AI scientist, including non-AI related articles. Here, the average number of papers by all AI scientists connected to a country is equal to 1. The size of the circles is equal to the share of the mobility category in the total number of AI scientists connected to a country. The figures of the other countries can be found in the appendix.  

If we look at the citation impact of AI scientists per mobility category, we see that this is generally higher for mobile scientists than for non-mobile scientists (y-axis). Mobile scientists also have on average a higher productivity (x-axis). The only exception is the US, where the citation impact score of non-mobile scientists is higher than that of departing scientists.

The relationship between mobility and production can partly be explained logically by experience: mobile scientists have been active as scientists for longer than non-mobile scientists (measured by the number of years they publish in Scopus). In general, there are relatively more mobile scientists among scientists with more publications (Robinson-Garcia et al., 2019). Moreover, previous research shows that when the number of publications of a scientist is higher, the citation impact scores of his or her articles are also higher (Larivière and Costas, 2016). At the same time, research by Robinson-Garcia et al (2019) shows that even when the number of publications is taken into account, mobile scientists have a higher citation impact than non-mobile scientists.

For the time being, we have no explanation for the difference in the US. There, too, non-mobile scientists have less experience on average. Nevertheless, their citation impact is higher than that of the departing scientists. We will discuss the US further on.

Figure 4 Quality and productivity of AI scientists per mobility category 

For almost all countries in our comparison, departing scientists have a higher citation impact than incoming scientists, as the first figure below shows. Exceptions are the US and the UK, where the citation impact of the incoming scientists is higher.

The average citation impact of scientists leaving the Netherlands (2.49) is 17% higher than that of incoming scientists (2.12). This is the smallest difference of all countries in this analysis with a negative ratio between the citation impact of the inflow and outflow. In Japan, the absolute difference is smaller, but because the citation impact scores are lower, it is slightly larger in percentage terms (23%). For Canada, Germany, Spain, Iran, and India and, the citation impact of departing AI scientists is more than 30% higher than that of incoming AI scientists.

The citation impact of scientists departing from the Netherlands (2.49) is, on an international level, average. It is comparable to the outflow scores from the US, France, the UK, and the EU-27. The Netherlands is also in the middle of the league with the citation impact score of incoming scientists.  

Like the UK, the Netherlands is a brain exchange country. Both countries combine a high overall quality of AI scientists with a high proportion of mobile AI scientists and a balance between inflow and outflow. In this way, these countries succeed in retaining many scientists with a high citation impact for a shorter or longer period of time. The citation impact scores of the different mobility categories are relatively close together. In France, too, the differences in the citation impact scores of the various mobility categories are relatively small, but the proportion of departing scientists here is 8 percentage points higher than the proportion of incoming scientists.

For the US, the average citation impact score of non-mobile scientists is remarkably high (2.88). For the other countries in this comparison, the citation impact of this group does not exceed 2.09 (UK). The citation impact score of scientists who spend only a short period of time outside the US is also remarkably high, at 3.01. This is followed by the Netherlands, at 2.15. The only mobility category in the US with a higher citation impact score is the incoming researchers (3.47). That of the long-term US departing researchers is 2.46. This makes the US the only brain gain country in our selection.

For Canada, the average citation impact score of the departing researchers is remarkably high. At 3.54, this is the highest average of all groups in all countries. This score is well above the average of all scientists associated with Canada (2.26). 

On the other hand, China has a very low mobility, combined with a low average citation impact score for non-mobile scientists. 

The Netherlands has a relatively large number of international co-publications

Not only are the AI-scientists associated with the Netherlands very mobile, they also collaborate a lot with foreign institutions. 60% of the Dutch AI publications identified for this study are the result of international colaboration. This is shown in Figure 5. For the UK, France, and Canada, the share of international co-publications in the field of AI is also above 50%. For China, only 24% of AI publications are the result of international colaboration.

In conclusion

The picture that emerges from this fact sheet is that the Netherlands is a brain exchange country. The Netherlands has a high circulation of AI-scientists and thus succeeds in connecting scientists with a high citation impact for a shorter or longer period of time. There is also a great deal of international colaboration.

The mobility of AI scientists therefore corresponds with the mobility of Dutch scientists in general, as can be seen in our research into the international mobility of scientists. 

This international mobility of scientists is seen as important at European, national and institutional levels. It is expected to contribute to the circulation of knowledge, increasing the network of scientists and their personal development (Koier et al., 2017). Scientists themselves also experience a positive effect on their research, network and personal development (Scholten, Koier and Horlings, 2017).  

At the same time, international mobility is also a cause for concern, especially in emerging scientific fields such as AI (Ministry of Education, Culture and Science, 2019). Where incoming talent can strengthen the economy, too much departing talent can weaken it.

Based on the results from this fact sheet, we can say that the international mobility of AI scientists is not at the expense of the quality of AI research in the Netherlands. To complete the picture, there is still a need for information about the international mobility of recently graduated AI-talent and of researchers from/to industry. At the moment, there are no sufficiently suitable quantitative, internationally comparable data available to gain insight into this.