Machine learning and the labour market: A portrait of occupational and worker inequities in Canada

INTRODUCTION

Industrialized labour market contexts are increasingly shaped by inequity which are known to influence disparities in the health and quality of life of workers¹. The rapid adoption of artificial intelligence (AI) has the potential to drive innovation and raise productivity. AI can also contribute to the polarization of the labour market. To date, most studies on the impact of AI focus on aggregate labour market outcomes. There is a lack of descriptive research which examines to examine how AI may differentially impact workers according to occupational or sociodemographic characteristics. Such research allows researchers and policymakers to identify groups of workers who may be most or least likely to be affected by the technology, and who may benefit from direct policy and programmatic support. We use a novel analytical approach and national labour force data to estimate the number of Canadian workers in occupations with high or low machine learning exposure. We then focus on describing the occupational and worker sociodemographic characteristics associated with high or low occupational exposure to ML.

Employment is a critical social determinant of health that provides income and resources that offer pathways to better health ¹. In Canada, like many other industrialized contexts, the labour market can be segregated according to occupational and sociodemographic factors ^{2 3}. Technological transformations have not only been a significant catalyst for economic change within the labour market but have also created both opportunities and barriers for different groups of workers. Previous of periods of technological transformation have increased the likelihood of labour substitution, contributing to partial or full job displacement for workers in certain occupations and potentially eroding working conditions for those who remain employed^4–6. Workers better positioned in the labour market, including those with more educational attainment and in better paying and higher-skilled jobs have been more likely to capitalize on economic advantage stemming from past technological transformations (e.g., wage growth and access to jobs with better quality). In comparison, workers from socially and economically disadvantaged groups have been adversely affected from past technological transformations and including experiencing job displacement or an erosion to job quality^{6 7}.

Research on the automation of work provides an example of how a technological transformation can widen inequities. Past periods of automation have shown that job tasks most likely to be performed by computer software, robots or other digital technologies included those that were routine, structured or repetitive that are most common in administrative, manufacturing or service-oriented occupations ^{5 6 8}. Research also indicates that worker sociodemographic and occupational characteristics were associated with the automation of work; workers with greater educational attainment and job skills, experience and training requirements and those that earn greater wages may be more likely to work in jobs with variable or cognitively demanding job tasks that were less susceptible to automation^9–11. Studies of past waves of automation have also indicated gender difference¹². Industries to which men are employed (e.g., repair, construction, and transportation) may be more likely to characterized by repetitive job tasks that are most suitable to being performed by robots or machines¹³ and the requirement for social and emotional job skills in occupations disproportionately held by women (e.g., education, healthcare, and administration) could be protective against automation ¹⁴. Other research suggests that women may be less likely to be employed in occupations with managerial or professional occupations or are underrepresented in science, technology, engineering and mathematics jobs that may benefit increased productivity related to the introduction of new digital technologies ^{12 15 16}.

Our study focuses on the implications of machine learning (ML), an AI subfield, for the employment and equity of the Canadian working population. ML involves the application of statistical algorithms to large amounts of data with the aim of making probabilistic inferences about a range of outcomes. ML is being adopted by workplaces to detect patterns in structured or unstructured data, make predictions, generate decisions, or optimize processes ^{17 18}. Examples of the use of ML exist across all industries, such as healthcare settings (e.g., use of ML to detect disease in medical imaging ¹⁹), financial services (e.g., ML-based fraud detection tools ²⁰) and the transportation sector (e.g., ML real-time hazard identification in automated vehicles²¹). ML may have a distinct impact on the labour market in ways that are different from past waves of automation ⁵. The capacity for ML to learn, adapt, and generate outputs with increasing independence means that the technology can perform job tasks that are physical or cognitive in nature and are across a broader range of occupations ¹⁷. It is important to highlight that rapid advancement in ML technologies and its processing power can double the speed of ML every 4 to 9 months, resulting in an evolving learning capacity of the technology and constantly changing impact on workers and workplaces ⁵. Innovations in deep learning, a subset of ML using neural networks (i.e., algorithms with structures inspired by human brain function) are also expected to enhance ML performance and its capacity to automate a greater range of job tasks including those that involve image and speech recognition and predictive analytics which could have significant implications for the working population.

Studies examining the extent to which job tasks that make up different occupations are suitable to being performed by ML (from this point forward referred to as occupational ML exposure) in the US and Canada suggest that almost all occupations include at least some job tasks that can be performed by ML ²². Other research indicates that 19 percent of US occupations may be highly exposed to ML ²³. Of note, mirroring past waves of automation, no single occupation can be fully performed by ML ⁵. As a result, it is expected that ML is unlikely to be used by workplaces to fully substitute human workers for machines but instead enhance the productivity of certain groups of workers ^{5 24}. Unlike previous waves of automation, ML doesn’t only target routine, repetitive and simple tasks, but can contribute to the automation of job tasks that can include planning, learning, reasoning, problem-solving or prediction that tend to cluster in occupations with greater job skills, training and experience requirements and among workers with greater levels of educational attainment ¹⁷. Accordingly, the impact of ML on the labour market is thought to have a unique and significant impact on occupations and different groups of workers. It is unclear whether occupational ML exposure which may be suitable to performing a greater range of job tasks to may exacerbate or minimize inequities in the Canadian working population.

Describing differences in the labour market related to occupational ML exposure represents an important strategy to build an evidence base that can be used to identify and address emerging occupational and worker inequities and examine how the rapid adoption of the technology can have implications for the social determinants of health. Our study describes divisions in Canada’s labour market related to the ongoing adoption of ML and help to identify groups of workers most affected by the technology. Our research has three main study objectives, which include:

1. Estimating the number of workers in occupations characterized by high or low ML exposure in Canada.

2. Examining how ML exposure differs according to workers’ socioeconomic (e.g., educational attainment, gender) and occupational characteristics (e.g., job skill level requirements, hourly wages).

3. Determining whether gender moderates the association between educational attainment, job skill requirements, hourly wages, and ML exposure.

METHODS

We conducted a pooled cross-sectional analysis of eight years of Statistics Canada’s Labour Force Survey (LFS; 2013-2019, 2022). The LFS is a nationally-representative cross-sectional monthly survey of the household population (≥15 years of age) with an approximate sample of 100,000 Canadians built using probability sampling procedures ²⁶. Excluded from the survey’s coverage are persons who are not currently employed, self-employed, living on reserves and other Indigenous settlements, full-time members of the Canadian Armed Forces, the institutionalized population, and households in extremely remote areas with very low population density. Weighted monthly cycles were combined to produce annual estimates. LFS waves 2020-2021 were excluded from the analysis due to the considerable economic impact of the COVID-19 pandemic which may have significantly skewed study results, given the unequal impacts of COVID-19 on labour market participation across industries and labour force subgroups ²⁷.

RESULTS

Overall, 1,902,050 Canadian workers were employed in occupations with high ML exposure where the greatest proportion of job tasks are suitable for ML representing (12 percent of the Canadian workforce). In comparison, about 744,250 workers were employed in occupations with low ML exposure where a smaller proportion of job tasks are suitable for ML (4.7 percent of the Canadian workforce). A full list of occupations categorized as high or low ML exposure is presented on Appendix 2. Women were more likely to be employed in occupations with high ML exposure than men (63.4 percent vs. 36.6 percent). Men were more likely to be employed in occupations with low ML exposure than women (59.9 percent vs. 40.1 percent) (Figure 1). The proportion of Canadian workers in occupations with high or low exposure to ML remained stable over the 8-year measurement period (Appendix 3).

• Download figure
• Open in new tab

Figure 1.

Number Canadian workers with high and low occupational exposure to machine learning (ML)

Table 1 summarizes the descriptive statistics for gender, educational attainment, hourly wages, and job skills, training, and experience requirements according to high or low occupational exposure to ML. Occupations characterized by high ML exposure were more likely to consist of a greater proportion of women workers when compared to men worker (63.4 percent vs. 36.6 percent). In contrast, occupations characterized by low ML exposure were more likely to consist of men compared to women (59.9 percent vs. 40.1 percent).

The proportion of workers in high and low exposure to ML also differed according to educational attainment. When compared to occupations with low ML exposure, occupations characterized by high exposure to ML were more likely to be composed of workers with a trade’s certification or bachelor’s degree (42.7 percent vs. 49.4 percent) and college or bachelor’s degree (38.6 percent vs. 34.6 percent). When compared to occupations with low ML exposure, occupations characterized by high exposure to ML were less likely to be composed of workers with some post-secondary education or less (8.2 percent vs. 4.6 percent) or a university degree above a bachelor’s degree (14.4 percent vs. 7.4 percent).

When compared to occupations with low exposure to ML, occupations with high exposure to ML were composed with a greater proportion of workers in hourly wage quartile 1 (21 percent vs. 29 percent) and hourly wage quartile 2 (19 percent vs. 33 percent). Workers with low exposure to ML were more likely to consist of workers in hourly wage quartile 4 when compared to occupations with high ML exposure (35 percent vs. 12 percent).

In comparison to occupations characterized by low ML exposure, a greater proportion of high ML exposure occupations were composed of workers in general skilled occupations with the lowest amount of job skills and training requirements (73.9% vs. 20.7%). Occupations characterized by low ML exposure were composed of workers in jobs with greater job skills and training requirements than those with high occupational ML exposure.

Separate gender-stratified multivariable models examined the association between educational attainment (models 1), hourly wages (models 2) and job skills, training, and experiences requirements (models 3) and the odds of high occupational ML exposure when compared to all other Canadian workers (a) or low occupational ML exposure when compared to all other Canadian workers (b) when controlling for study covariates (Table 2). Statistically significant gender differences were identified in all models and key findings will be described in the sections below.

For both women and men workers, greater educational attainment was associated with a lower odds of high occupational ML exposure when compared to Canadian working population and when controlling for study covariate (model 1a). Gender-specific effects emerged when examining odds of low occupational ML exposure (model 1b). Women with a university degree above a bachelor’s degree (OR = 1.64, 95%CI, 1.59-1.69), college or bachelor’s degree (OR = 1.27, 95%CI 1.24-1.29) or trades certification or diploma (OR=2.72, 95%CI, 2.21-2.34) had a greater odds of low occupational ML exposure when compared to women with some post-secondary education or less (lowest educational attainment category). Men with a university degree above a bachelor’s degree (OR = 0.80, 95%CI, 0.78-0.82) or a college or bachelor’s degree (OR = 0.92, 95%CI, 0.91-0.94) had a lower likelihood of low ML occupational exposure when compared to men with lowest educational attainment.

• Download figure
• Open in new tab

Figure 2.

Summary of multivariable logistic regression models examining the association between educational attainment and high occupational machine learning (ML) exposure (a) and low occupational ML exposure (b).

Among both women and men, the highest earners were less likely to make up the workers with high occupational ML exposure (model 2a) and more likely to make up the workers in low ML exposure occupations when controlling for study covariates (model 2b). Of note, women with hourly wages in the 2^nd quartile (OR = 1.85, 95%CI 1.83-1.87) and 3^rd quartile (OR = 1.56, 95%CI 1.54-1.58) had a greater likelihood of high occupational exposure to ML when compared to lowest wage women earners (quartile 1).

• Download figure
• Open in new tab

Figure 3.

Summary of multivariable logistic regression models examining the relationship between hourly wages and high occupational machine learning (ML) exposure (a) and low occupational ML exposure (b).

Finally, greater job skills, training and experience requirements were associated with a lower likelihood of a high occupational exposure to ML (model 3a) and higher likelihood of reporting low occupational exposure to ML (model 3b). Interestingly, women in managerial positions (with the highest job skills, training, and experience requirements) had a greater odds of low ML exposure when compared to women workers in general skilled occupations (OR = 1.77, 95CI% 1.70-1.83). Men workers in managerial occupations were less likely to be low ML exposure when compared to those with general skilled occupations (OR = 0.86, 95%CI 0.83-0.89).

• Download figure
• Open in new tab

Figure 4.

Summary of multivariable logistic regression models examining the relationship between occupational job skills and training requirements and high occupational machine learning (ML) exposure (a) and low occupational ML exposure (b).

DISCUSSION

The adoption of ML within Canadian workplaces has the potential to drive rapid change to the nature and availability of work and may also exacerbate workforce inequities that can disrupt access to critical social determinants of health. Our study is among the first we are aware of that provides a snapshot of occupational ML exposure in Canada’s labour market and the extent to which high and low occupational ML exposure may be associated with worker sociodemographic and occupational characteristics. Like other technological transformations that have shaped Canada’s labour market, we show that historically disadvantaged segments of the workforce may be most likely to have their occupations affected by ML. We also show that ML may have a gendered effect and disproportionately impact women when compared to men. Findings can inform targeted policies and programs that optimize the benefits of ML and address the potential adverse effects of the technology on workers, especially amongst groups with the highest occupational exposure.

We utilized a novel analytical approach to describe Canadian occupations based on the extent to which they are made up of job tasks that are suitable to being performed by ML. Through this approach we estimated the number of workers in occupations with the highest (i.e., occupations most suitable to being performed by ML) and lowest exposure to ML (i.e., occupations least suitable to being performed by ML). Through our analytical approach we show that close to two million Canadian workers are employed in occupations with high exposure to ML, which represents close to 11.5 percent of the overall workforce. Less than half of that number of Canadian workers were in occupations characterized by low ML exposure (744,250 workers representing 4.5 percent of the Canadian workforce). It is not clear based on our analytical approach whether workers in jobs with high exposure to ML will experience advantage or disadvantage stemming from the technology. Drawing from past research there are several ways that our findings could be interpreted. First, workers in occupations with high exposure to ML could be at risk of partial or full job displacement. Second, the use of ML may complement the work performed by workers in high exposure jobs and free up time for high value activities and productivity gains ^{4 32 33}. On the other hand, while those with low occupational ML exposure are less likely to work in a job where they face less of a risk of being displaced by a machine, they may also be less likely to leverage benefits associated with ML to the performance of their job tasks. Additional research is needed to expand on our results and specifically examine how ML both positively and negatively impacts the employment and working conditions for different groups of workers.

We also found that all Canadian occupations consisted of job tasks that could be performed by ML. Findings align with past studies in Canada and the US which indicate that the technology can have a widespread impact on the economy ^{4 22}. At the same time, our study show that no single occupation could be completely suitable to being performed by ML and the ability for machines to match humans in all job tasks remains limited ⁵. Findings support an ongoing perspective that ML will contribute to changes to Canadian workplaces and the role of workers. There is a need to prepare all workers for the growing adoption of ML and other AI so that they can work alongside technology. The importance of studying the impact of ML on the nature of work will continue to grow as the technology advances and gains greater independence and surpasses human learning and reasoning ^34–36.

Scholars have previously posited that the adoption of ML within the labour will have a disproportionate impact on workers in higher-skilled and better paying occupations whose jobs tasks require a greater degree of prediction ^{9 17 37}. Our results, however, align with research of past waves of automation and show high occupational ML exposure among workers with less educational attainment or those in lower-paying occupations and occupations with fewer job training, skill, and experience requirements. Our study signals a potential risk that the adoption of ML could pose to the labour market by reinforcing or even widening inequities. Of particular concern, ML has the potential to contribute to wage polarization between groups of workers with high and low exposure to ML. The transfer of work responsibilities from workers to machines is expected to lower the cost of labour and contribute to a downward pressure on wages which could disadvantage vulnerable worker groups ^{4 37}. Alternatively, ML may be used to perform the most repetitive or unsafe job tasks and contribute to working conditions that are more optimal to groups of ^{38 39}. Additional research is needed to elaborate on our study findings and build a better understanding of the mechanisms by which ML may positively or negatively impact working conditions and different groups of workers to develop targeted policy interventions and supports.

Occupational exposure to ML in the Canadian labour market may also be patterned according to gender. Men and women tend to cluster in different occupations which shape their employment experiences and exposure to different technologies in the workplace ^{13 40}. Our findings add to the evidence on gender segregation in the labour market related to technological transformations. We show that Canadian women may be employed in occupations where their tasks are most suitable to ML and may be more likely to experience high exposure to ML ³. The implications of these findings could be interpreted in several ways. The job tasks that women workers perform and the job roles they hold may be at a greater risk of being substituted by ML. Alternatively, women may also be more likely to benefit from ML including the technology’s ability to complement their tasks and contribute to productivity gains. At the same time, our findings may suggest that men may be shielded from job displacement related to ML. Men may also be more likely to be excluded from the economic opportunities that can emerge because of ML adoption. Findings highlight the importance of future research to unpack how ML adoption may impact men and women differently and the potential need for gender-sensitive policy and programmatic approaches to address the challenges and opportunities of ML for Canadian workers.

To further unpack gender differences, our study examined whether the relationship between educational attainment, hourly wages and job skills, training and experience requirements and ML exposure differed for women and men. Interestingly, our findings indicated that greater educational attainment and being in a managerial occupation was associated with low occupational ML exposure among women. Past studies have indicate that educational attainment may be particularly important among women and insulate them from economic challenges including automation risk ^{9 41}. At the same time, our study could suggest that highly educated women and those in management positions may be excluded from ML-related economic opportunities that could reinforce or widen gender divides. Future research is required to better understand the how ML may impact the employment experiences of men and women in the labour market and how differences may stem from worker and occupational characteristics.

Our study has several strengths and limitations. We use a novel analytical approach that measures the extent to which different Canadian occupations are made up of job tasks that are suitable to being performed by ML. We use eight cycles of Canada’s LFS to produce a population-based estimates of ML exposure. We were also able to describe occupations with high and low ML exposure according to occupational and worker characteristics. Findings provide an important description of the extent to which ML may impact Canada’s labour market and the ways in which the technology could contribute to labour market segregation. There are also some limitations to note. Our measure focuses on the technical feasibility of ML and its suitability to perform job tasks ^{4 22}. We are unable to comment on how ML is used in practice and the specific ways in which it will impact workers and the way they perform their jobs. Our findings do not elaborate on the economic, organizational, legal, cultural, and societal factors that may facilitate or inhibit ML adoption within the labour market and its impact on occupations and workers.

While we utilize an established measure of occupational ML exposure, other measures exist including those that capture AI exposure at the industry or geographic levels ⁴² as well those which examine the extent to which AI may be used to complement or substitute workers ³². Moreover, the measure we selected does not evaluate all forms of AI and the automation of work such as those driven by the most recent advances in generative AI ²⁵. Recent measures have been developed to examine the occupational impact of generative AI ²⁵. These measures are preliminary and require further validation before being more widely used for labour market estimates. Consequently, ongoing research is required to examine how emerging forms of AI will impact workers and workplaces.

CONCLUSION

ML, an AI subfield, has the potential to transform Canada’s labour market and contribute to labour market and health inequities. Our analysis of labour force data offers a critical picture of the occupations most affected by ML and indicate that a substantial number of Canadians are working in jobs where they may be likely to be exposed to ML. Workers in jobs with high exposure to ML may face both challenges and opportunities associated with the technology. We also highlight that occupational ML exposure may contribute to the segregation of the labour market according to educational attainment, wages and occupational job skills and training requirements. Our study provides a foundational evidence base that can be used to direct attention and efforts towards occupations and worker groups who may be most likely to directly be exposed to the ML to develop strategic responses that ensure inclusion in a working world where ML and other forms of AI is being increasingly used.

FUNDING STATEMENT

The study was supported by Social Sciences and Humanities Research Council of Canada Partnership Development Grant (#890-2021-0018), and the Future Skills Centre (#20220524). Dr. Arif Jetha’s salary is partially supported by a Stars Career Development Award from the Arthritis Society (Canada). Funding bodies had no role in study design or manuscript writing.

COMPETING INTERESTS

AJ, QL, FVS, AB and PS are employees of the Institute for Work & Health (IWH), which is supported through funding from the Ontario Ministry of Labour, Immigration, Training and Skills Development (MLITSD). The analyses, conclusions, opinions, and statements expressed herein are solely those of the authors and do not reflect those of the MLITSD; no endorsement is intended or should be inferred.