The geography of job automation in Ireland: what urban areas are most at risk?

Humans have always developed new and superior products and production technologies to satisfy new wants and needs and to produce greater economic output with less human effort. This is what is meant broadly by technological change, and it is a significant catalyst of long run economic growth (Romer 1990; Schumpeter 1942). In this context, the idea of computerisation, where job automation is by means of computer-controlled equipment (Frey and Osborne, 2017), is not new. New technologies (like computerisation) and new markets have replaced and generated new and more productive jobs and will continue to do so in the future (OECD 2018). Automation technologies have a long history, with inventions like the steam engine, electricity, mechanised weaving looms, industrial robots, or automated teller machines and more recently with information technologies (Acemoglu and Restrepo 2018). From this perspective, robots have been around for a long time.

As far back as Simon (1965), the decline of routine jobs was predicted, where computers were argued to hold the comparative advantage in ‘routine’ rule-based activities which are easy to specify in computer code. Autor et al. (2003) presented a simple model, referred to as the ALM model, of how the rapid adoption of computerisation by firms is changing the tasks performed by workers at their jobs, which in turn changes the market demand for human skills. For them, robots (machines) are substituting ‘routine’ tasks but will not substitute non-routine tasks that involve problem-solving, complex social and emotional communication activities and tacit knowledge. Non-routine tasks that cannot be substituted by automation are generally complemented by it (Autor 2015). Routine tasks are tasks that can follow explicit computerised code rules, and by way of simple contrast, non-routine tasks cannot be specified in a computer code (Frey and Osborne 2017). Autor (2015) argues that the tasks that can be automatable are bounded by the limits of human knowledge and what he refers to as Polanyi’s paradox: Since humans ‘know more than they know they know’ (i.e. know things that are difficult to explain as a matter of codified, programmable steps), there are limits to the substitutability of human tasks. Consequently, there will always be a division of ‘human tasks’ and ‘machine tasks’.

Susskind and Susskind (2015) have identified key weaknesses of the theoretical ALM model, where tasks that were considered ‘non-routine’ in the ALM task-based literature have already become ‘routine’, due to technological changes leading to the likelihood of a more dramatic job displacement reality, from current modern computerised technologies. Big data and artificial intelligence are progressing at incredible speed. Susskind and Susskind (2015) observations pre-empted the findings of Frey and Osborne (2017), who identified the scope of jobs that could be automatable in the future. Frey and Osborne (2017) identified that 47% of jobs in the USA are susceptible to automation, at some unspecified time in the future. They argue their estimate also aligns closely to the extent of job displacement that has happened in the past. For instance, in 1900, 40% of the US workforce was employed in agriculture. Now it is less than 2%.

In general, the grave concern of technological unemployment from theorists such as Keynes (2010) has been unfounded. The process of creative destruction has led to as many new jobs being created in new sectors, as destroyed in traditional sectors. Hence, jobs lost were replaced, with often better paid, more highly skilled jobs in their place. Despite the unfounded unemployment concerns of the past, automation still created widespread disruption with a greater polarisation in jobs and incomes, as technologies (robots and algorithms) replaced, middle-skilled (in particular) people based-routines (OECD 2018). And future disruption is predicted. Schwab (2016) reflecting on the fourth industrial revolution, at the economic summit in Davos stated ‘We stand on the brink of a technological revolution that will fundamentally alter the way we live, work, and relate to one another. In its scale, scope, and complexity, the transformation will be unlike anything humankind has experienced before’.

Job polarisation and technological unemployment are often the most discussed adjustment concerns raised around automation and the rise of artificial intelligence. But an often-overlooked implication and, possibly a greater problem, is the displacement to places created by technological change. The past century has resulted in a large migration of people from rural to urban areas. By 2050, Eurostat (2016) predicts that about 80 per cent of the European population will be living in urban areas.

Urban theorists argue that agglomerations are intellectual breeding grounds for new ideas and innovations (Jacobs 1969; Marshall 1890). The geographical proximity between economic actors helps foster, facilitate and nurture face to face interaction, the flow of local knowledge, ideas and innovations (McCann and Shefer 2005: 302). In theoretical approaches to knowledge flows and innovation, it is assumed that it is easier for knowledge to transfer over shorter distances than longer distances, or as Glaeser et al. (1992: 1126) put it, ‘intellectual breakthroughs must cross hallways and streets more easily than oceans and continents’. Given the spikey distribution of prosperity around the world, between countries and within countries, the importance of ‘place’ in contributing to economic growth has been a central research topic (Florida 2005).

From the literature, it is apparent that urban areas play a critical role as creative and intellectual spaces (Frank et al. 2018). The ‘tolerant’ and ‘bohemian’ nature (Florida 2005); the creative spill-overs that are ‘in the air’ (Marshall 1890); dense diversity (Jacobs 1969); and the creative ‘buzz’ (Storper and Venables 2004) in urban areas drive innovation resulting in ‘swarms of creative clusters’ to emerge in global and capital cities (Chapain et al. 2010). Florida’s thesis outlines that the disparity in the creative class across cities is driving the disparity in innovation and productivity outcomes across space. Educated, creative, more highly skilled people are migrating to urban areas to exploit amenities, higher wages and employment opportunities in favourable business environments (Moretti 2012; Florida et al. 2011; Marlet and van Woerkens 2007). Patterns of human capital accumulation are following a self-propelling process, where well-educated individuals are attracted to areas that already are concentrated with well-educated individuals (Waldorf 2009). Consequently, ‘superstar’ agglomerations are providing an enhanced ecosystem to drive novel combinations of knowledge and ideas which are further compounding previous geographical divides between regions (Iammarino et al. 2018) and urban and rural areas (Moretti 2012; Rosenthal and Strange 2004) through time.

The shift from manufacturing to predominantly service driven economies has made geographical proximity matter more, resulting in increased divergence between places (Moretti 2012). There is also strong evidence that digital technologies are resulting in greater divisions between large and small urban areas (Muro 2019; Frank et al. 2018). In general, urban areas with a more critical mass are more productive (Melo et al. 2017; Coulibaly et al. 2008), more innovative (Carlino et al. 2007), more accessible to industries and anchor institutions (Hamidi and Zandiatashbar 2019) offer more job opportunities (Hendrickson et al. 2018) and higher wages (Wang 2016; Echeverri-Carroll and Ayala 2011).

But it is also not just a story of ‘large’ versus ‘small’ and ‘leaders’ versus ‘laggards’—a much more complex geographical landscape is emerging with brain hubs, declining manufacturing cities and a number of cities that could thrive or decline (Moretti 2012). The reality of shrinking urban areas is not a recent phenomenon (O’Driscoll et al. 2022; Wolff and Wiechmann 2018). It is suggested that economic stagnation resulting from a decline of economic and social opportunities in smaller cities, declining regions and rural areas, was the force behind recent populist voting patterns in the USA (Hendrickson et al. 2018; Frey et al. 2017; Shearer 2016), in France, Germany and the UK (Rodríguez-Pose 2018; Dijkstra et al. 2018; Ballas et al. 2017). In response, many are calling for new frameworks, strategies and policies to tackle the concerns of lagging areas (Rodríguez-Pose 2018; Shambaugh and Nunn 2018; Storper 2018; Iammarino et al. 2018).

This regional disparity may be exacerbated and intensified by an uneven spread of automation risk. Indeed, the aggregate number of jobs at high risk of automation at a national level hides the sub-national regional variation of job exposure. Automation is predicted to have a different impact across regions (OECD 2019). The share of jobs at high risk of automation across OECD defined territorial level 2 regions varies considerably from a low of 4% to a high of 40% (OECD 2018). It is also predicted that automation will result in improved welfare (particularly for better educated workers) in leading regions, and losses in welfare (particularly for lower educated workers) in lagging regions, as production will shift to locations with higher wages and skills, due to more efficient capital over time (Lecca et al. 2018). Indeed, Okamoto (2019) identified that individuals relocate to states with more automation in skilled occupations and less automation in unskilled occupations, demonstrating that automation has a complementary effect on skilled occupations and a substitution effect on unskilled occupations. In effect, if the adoption of present and future automated technologies is rapid, and new firms in new technological industries follow the similar past trends of high knowledge-based firms, by spatial clustering, by locating around successful cities, talent and centres of learning and research (Florida et al. 2011; Venables 2010; Berry and Glaeser 2005; Florida 2005), then it is inevitable that future automation disruptions will further exacerbate regional and local disparities. Present trends in the area of artificial intelligence suggest that the industry is already following a clustering and localised co-location pattern (Klinger et al. 2018). Barzotto and De Propris (2021) identified that firms which collaborate with local suppliers of enabling technologies linked to industry 4.0 have improved levels of performance.

So far, the extent of the automation job risk literature has focused on the level of national exposure (Frey and Osborne 2017), European regional territorial level 2 exposure (Crowley et al. 2021; OECD 2018; Arntz et al. 2016) or US county and city-level exposure (Muro 2019; Frank et al. 2018). While the OECD (2018) and Crowley et al. (2021) papers are significant steps in assessing the job exposure to automation across different regions, such (regional territorial level 2) defined boundaries are unlikely to identify deeper local labour market areas in the European case (Coombes 1995). It is common across countries for development plans and policies to transcend across many different levels of government, and it is typical to have a combination of local, regional and/or national intervention programmes for economic development (OECD 2016a, 2018). Consequently, it is important to have a competent empirical understanding of local labour market dynamics (Leigh and Kraft 2018). This paper focuses specifically on ensuring a deeper disaggregated regionalisation of automation job exposure to urban settlements in Ireland to get a better understanding of the localised automated job risk exposure across Ireland.

Within O*NET and the automation literature, it is common to reference occupations, jobs, and tasks. Here we provide a summary of what these mean in the context of our paper. Every occupation requires a different mix of knowledge, skills, and abilities, and is performed using a variety of activities and tasks (ONET Online 2022a). It is these activities and tasks which define an occupation with each occupation involving a different mix of tasks (US Department of Labor 2022). Tasks are defined within O*NET as the specific work activities undertaken within an occupation and are identified as the critical tasks associated with each occupation in the database (ONET Online 2022b; Van Iddekinge et al. 2003). These tasks are the day-to-day aspects of a job undertaken by a worker. As an example of the distinction between a job, an occupation, and a task, Appendix 1 presents a breakdown of the jobs and tasks association with the ‘Construction Laborers’ occupation. Within this occupation, job titles may include ‘construction laborers’, ‘construction workers’, ‘drop crew laborers’, amongst others listed within the appendix. While the tasks that these workers perform, within this occupations, include the ‘read[ing of] plans, instructions, or specifications to determine work activities’ or the ‘lubricat[ion], clean[ing], or repair [of] machinery, equipment, or tools’, amongst others listed within the appendix.

Frey and Osborne (2017) (hereafter referred to as FO) obtained data on the working population of the USA from O*NET surveys. O*NET data contain approximately 20,000 unique task descriptions and data on the skills, knowledge, and abilities of different occupations in the workforce. FO created a ‘training dataset’ using machine learning experts to assess the automatability of 70 occupations using detailed task descriptions. Specifically, they asked the experts to assess whether each task for these occupations was likely to be automated given current knowledge on computerisation capabilities and possibilities. Nine properties were used to assess occupations such as the level of manual dexterity or social perceptiveness.¹ They then used AI algorithms from the training dataset to examine and create an automation probability for a total of 702 occupations that correspond to the 2010 version of the US Standard Occupational Codes (SOC), making up 97% of the U.S. workforce. We base our analysis on transposing these 702 occupations to the Irish context.

The FO approach to measuring automation has proved contentious. This contention has primarily arisen due to the underlying assumption in their method that entire occupations are at risk of substitution, despite the possibility that tasks may vary within an occupation, at the individual level, across different workplaces and locations (White et al. 2019; Frey and Osborne 2018; Nedelkoska and Quintini 2018; Arntz et al. 2016). To overcome these concerns, the task-based approach adopted by Arntz et al. (2016) and Nedelkoska and Quintini (2018) (hereafter referred to as NQ) uses the OECD’s Survey of Adults Skills (PIACC) which contains detailed worker data on how individuals in the same occupations may have different tasks across different workplaces, as well as demographic variables such as gender, education, age and income of workers. For example, construction labourers may have different tasks depending on the type of work being conducted and their individual capabilities, which may vary across different workplaces. The Arntz et al. (2016) study exploits the demographic data where for example an individual’s income or education level may increase or reduce their automation risk. Frey and Osborne (2018) criticise this approach asking the question ‘why should automation discriminate on the basis of worker characteristics?’ The NQ approach did not use worker characteristics and instead applied the FO automation probabilities to the task make-up of occupations at the individual level. NQ created a training data set using a logistic regression and data from the Survey of Adults Skills (PIACC) for Canada, as Canada had a substantially larger sample size than other countries. Firstly, they manually compared the 70 O*NET occupations used by FO with ISCO data. Secondly, they identified occupation task variables in PIACC that best corresponded to the nine engineering bottlenecks identified in FO.² Lastly, PIACC has individual level data on occupation tasks allowing NQ to control for within occupation differences that may occur at an individual level across different workplaces, which is the key differentiating factor between the FO and NQ approach. The estimated coefficients from the Canadian training set were then applied to all other individuals across 32 countries in PIACC. Frey and Osborne (2018) criticise the NQ paper for not providing greater clarity on the degree of variability of tasks within occupations, across different workplaces, and question why tasks performed by different workers within occupations would vary that greatly in the first place.

Widely contrasting predictions emerge when the two methods are applied to the entire labour market to predict the proportion of jobs at risk to automation. FO estimated that 47% of jobs in the USA were at high risk of being automated, whereas NQ estimated that 14% of OECD jobs are at high risk. However, NQ also estimated that a further 32% of OECD jobs have a 50–70% likelihood of significant change and the median job is estimated to have a 48% probability of being automated. The key distinction between the occupational versus task-based methodologies lies in the distribution of jobs from low to high risk of automatability. The FO method produces a U-Shaped distribution with extremes at both ends, whereas the Arntz et al. (2016) method produces an inverted U-shaped distribution. The NQ method produces similar inverted U-shaped distributions for countries across the OECD which can be seen in Nedelkoska and Quintini (2018). Despite the differential distributions patterns, what is consistent with both methodologies is the large proportion of jobs that is expected to experience significant change.

In our analysis, as a starting point, we use the probabilities of automation estimated by Frey and Osborne (2017), for 702 occupations in the US case, to identify the risk probabilities associated with occupational classifications in the Irish case. FO’s occupational classifications are based on the 2010 version of the US Standard Occupational Codes (SOC). The Irish Central Statistics Office (CSO) bases their occupational classifications on the UK SOC. The US and UK SOC are not directly comparable, and there is no direct conversion available. Therefore, in order to convert the US codes to their UK counterparts (which are approximately identical to the Irish codes used by the CSO) we transform these data using a series of established international classifications. This allows us to perform an analysis of the risk of automation to occupations in Ireland. Once this process has been completed our analysis begins with, what the CSO define as, the detailed occupational classifications for Ireland of which we have FO risk probabilities associated with 273 detailed occupations across 24 intermediate occupation aggregates, in total covering approximately 86% of all jobs in Ireland. Appendix 2 provides a more detailed discussion of the matching procedure and the occupations that cannot be considered in the Irish case.

The most disaggregated Irish data in terms of the scope of occupations are available at a national level. This is referred to by the CSO as the detailed occupations classifications and covers, in our analysis, 273 occupational classifications. This detailed classification is also available at Local Authority level of which there are 31 in Ireland. At these levels, data are available from the Irish censuses of 2016 on the number of people employed in each detailed occupational classification. Therefore, the exposure of Ireland, and the Local Authorities to automation, can be obtained by identifying the number of people employed across different ‘probability of an occupation being automated’ groupings. Maintaining alignment with the literature (OECD 2018; Frank et al. 2018; Frey and Osborne 2017; Arntz et al. 2016), we identify jobs at high risk of automation as those with a probability of automation greater than 70%. We further define medium risk as being 50 to 70% probability of automation and low risk as less than 50% probability of automation to align with the methodology employed in Nedelkoska and Quintini (2018).³

At a national level, using FO, we observe a general U-shaped pattern where jobs in Ireland are mainly susceptible to high (44.6%) or low (46.4%) risk of automation with relatively fewer jobs falling into the medium risk category (9%). This U-shaped pattern is largely in line with the findings of FO for the USA. At a national level, using NQ, 18.4% of jobs are at high risk, 8% are at medium risk, and the remainder are at low risk. When our analysis is compared with the NQ analysis of the OECD, the proportion of jobs at high risk is greater in Ireland, but a considerably lower proportion is at medium risk (32% of jobs in OECD are at medium risk).

In the Irish case, occupational classifications for towns with a population of more than 1500 people are only available at the intermediate occupational classification, of which there are 24 occupational classifications (see Appendix 3 for a detailed discussion of them). The lack of availability of detailed occupational classification is not limited to Ireland as it is common across most countries.

The regionalisation methodology applied in this paper of the FO approach is based on firstly identifying the automation risk for each detailed occupational classification nationally. The method of identifying the automation risk at town level can be summarised as follows (with a detailed example provided in Appendix 4):

The process for regionalising the NQ approach is more straightforward. The authors were provided by NQ with the proportion of the workforce at high and significant risk of automation at ISOC 2 digit level. This was mapped to the Irish intermediate SOC classifications, and then the number of people in each town employed in each intermediate occupational class was multiplied by the NQ proportions at high and significant risk of automation.

Therefore, at the Irish town level, all analysis is conducted based on aggregate intermediate occupational classifications regardless of whether we consider the FO or NQ approach. More routine-based occupations are at greater risk and for a breakdown of risk by occupation across the intermediate occupational classifications, see Appendix 5. We conduct checks to ensure our regionalisation approach is robust (which is outlined in Appendix 6 for interested readers).

In total, we examine 200 Irish towns (including Ireland’s five cities: Dublin, Cork, Limerick, Galway and Waterford) with a population over 1500 people. A map showing the distribution of these towns across Ireland and the spatial autocorrelation of automation risk at town level is presented in Appendix 7. The average percentage of jobs at high risk of automation across Irish towns using the FO approach is 44%, compared to 18% using the NQ approach, highlighting a large gap in average exposures between the different methodologies, which is consistent with this literature (Frey and Osborne 2018). However, the correlation between the FO and NQ proportion of jobs at high risk across towns is 0.9524 (with an associated p-value of 0.0000). The key observed difference is the downward shift in risk exposure associated with the NQ approach, relative to the FO approach.

Separately, a spatial dependence Moran’s I plot is estimated to identify significant town risk clusters of the proportion of jobs at high risk of automation, using the FO method. The results of the local spatial autocorrelation plot are presented in Map A and B in Fig. 1. Significant clusters (Map B) of at lower risk towns can be mainly found in the greater Dublin city region (dark blue circles), and separately, significant clusters of at higher risk towns (light green circles), can be found in the West and South-West of Ireland. More detail on this analysis is provided in Appendix 7.

We deepen our analysis by considering some factors which may determine the proportion of jobs at high risk of automation. We estimate Eq. (1) for the 200 Irish towns for which we have generated job risk data.where \({\mathrm{Automation} \mathrm{Risk}}_{i}\) is the proportion of jobs in town i at high risk of automation. The \(\mathrm{Independent} \mathrm{Variables}\) include: urban population size dummy variables as per Table 1; the age profile of town i dummy variables; the proportion of the population in town i aged 24 or over with a third level qualification; the proportion of the population employed in a given sector in town i; the proportion of the population in town i that are Irish; and the employment rate calculated as the number of people who indicate that they are employed, divided by the number of people who indicate that they are employed, plus the number of people unemployed, plus the number of people seeking their first job. The data are derived from the Irish Census 2016. Table A8.1 in Appendix 8 provides descriptive statistics of all variables utilised. Appendix 9 presents a variance inflation factor (VIF) table from the estimation of Eq. (1) to test for multicollinearity. The mean VIF is below 5 indicating that multicollinearity isn’t an issue.⁵ Appendix 10 presents a correlation matrix of our variables. We estimated a number of models as part of a sensitivity analysis including variations of estimation techniques and variations of the dependent and independent variables. These sensitivity analyses are outlined in Appendix 11 for interested readers, and the results are consistent and robust across different model specifications. Moving forward, we report the results from the ordinary least squares regression.