While most OECD economies are still in the early phases of AI adoption, the technology is expected to have a profound impact on the world of work. This chapter explores the evidence from across the OECD as well as from Korea on how AI has so far affected: job quantity and skills, job quality, and inclusiveness in the labour market. So far, there is little evidence of a negative impact on the number of jobs, although some groups are more affected than others. AI holds promise to improve job quality, but there are risks too. Moreover, these risks and benefits of AI are not equally distributed across population sub-groups.
Artificial Intelligence and the Labour Market in Korea
Abstract
The impact of AI on job quantity and skills
AI has made the most progress in non-routine, cognitive tasks. Therefore, the occupations most exposed to AI tend to be white‑collar occupations, such as IT professionals, business professionals, managers, and science and engineering professionals. However, high exposure to AI does not necessarily imply workers in these occupations will be displaced. So far, across OECD countries, there is little evidence of negative aggregate employment outcomes due to AI.
In Korea, there is some evidence that over the period 2018 to 2023, more “traditional” AI was associated with lower growth in full-time, permanent jobs, particularly in the manufacturing sector. However, no such relationship was found for generative AI. These findings need to be interpreted in a context of 2.4% employment growth in full-time, permanent jobs overall during the same period, as well as 5.8% growth in total employment (including non-standard forms of work). In addition, in a survey of Korean firms, 95.5% report no workforce changes so far at the department- or team-level following the adoption of AI.
These limited effects on job automation suggest that, by itself, AI will not solve labour shortages in Korea. However, AI can help mitigate them. For instance, of the 37% Korean SMEs reporting a worker shortage in the last two years, 27% say that generative AI helps compensate for these shortages. Similarly, 24% of Korean SMEs report a lack of skills and experience among staff, and 47% of these say that generative AI helps to address this challenge.
While AI may not automate jobs at large scale, it does change the tasks workers do and the skills required of them. 56.5% of Korean firms that have adopted AI say it has replaced specific tasks within existing jobs. Moreover, 32.2% say the use of AI has resulted in an increase in the kinds of skills required to carry out current tasks, and 38.3% say that AI has increased the level of skills required. Firms in Korea that have adopted AI are more likely to report an increase in communication among team members (18.4% v. 7.1%), percentage), with managers (19.1% v. 7.1%), as well as between teams (27% v. 7.8%). For Korean SMEs, AI increases the importance of data analysis and interpretation skills increases the most, followed by programming and coding skills.
The changing skills needs resulting from the adoption of AI in the workplace call for new training opportunities. This is particularly important in the context of the continued brain train from Korea, including of AI talent. However, Participation in adult learning in Korea is the lowest across OECD countries: 13%, compared to an OECD average of 40%. While firms in Korea do provide training to employees for working with AI, only 42% of those that have adopted AI have done so, and the share is higher in large firms than it is in small ones.
The impact of AI on job quality
Wages are a key dimension of job quality. If AI boosts productivity then it could result in higher wages for workers. In OECD countries, the wage benefits of AI have so far been concentrated among high-income and highly skilled workers. Similarly, in Korea, only the occupations most exposed to generative AI have benefited from higher wage growth.
AI could improve job quality in other ways as well. The automation of tedious and repetitive tasks could improve job enjoyment and allow workers to focus on more complex and interesting tasks. AI could also improve physical safety by automating dangerous tasks and improving monitoring systems and safety procedure controls. At the same time, there are some risks too. For example, case study evidence from OECD countries shows instances of increased work intensity due to higher performance targets or complexity induced by AI. Ultimately, the effect of AI on the work environment depends on how thoughtfully and strategically it is integrated into workplace practices.
In Korea, AI appears to improve job satisfaction. However, there is a gap between the perceptions of firms and employees, with the former being more positive than the latter. In addition, many workers in Korea – particularly those in smaller firms and the manufacturing sector – report no noticeable reduction in either physical or mental burden following the adoption of AI. One possible explanation is that AI adoption in Korea is still in its early stages, and its potential to ease work intensity has not yet fully materialised.
The impact of AI on inclusiveness
Workers vary in the extent to which they are exposed to AI, but also in their ability to adapt to and benefit from new technologies. Thus the impact of AI need not be uniform across different socio-demographic groups. So far, the evidence from OECD countries suggests that high-income and high-skilled workers benefit the most from AI, while low-skilled workers may lose out. For example, the impact of AI on employment growth has been found to be significant and positive for high-income and high-skilled occupations, and for jobs where computer use is high. Similarly, high-income and high-skilled occupations, as well as jobs with high computer use, tend to experience positive effects on wage growth associated with AI exposure, while lower income and lower skilled workers do not seem to benefit in the same way, or less so.
In Korea, the negative impact of traditional AI on regular, full-time employment growth appears to be concentrated among younger workers, low- to medium-skilled workers and women – although for the latter, as well as for high-skilled workers, higher exposure to generative AI is associated with higher employment growth. Furthermore, generative AI is associated with higher wage growth for men and high-skilled workers, while traditional AI is associated with higher wage growth for older workers and high-skilled workers. By contrast, traditional AI appears to reduce wage growth for low-skilled workers. These findings apply to full-time, permanent employees only, and the findings for non-standard workers may be different.
AI has made the most progress in non-routine cognitive tasks, therefore affecting mostly white‑collar occupations
Recent advances in AI have extended the types of tasks that can be automated to non-routine, cognitive tasks, exposing workers who were previously relatively protected from the risk of automation (e.g. the high-skilled). In the past, computers and robots followed strict rules set by programmers and therefore could only automate routine tasks, affecting mostly low- and medium-skilled workers (Autor, Levy and Murnane, 2003[1]).
As of 2023, AI has exceeded human performance across various tasks. It outperformed human baselines in image classification1 as early as 2015, basic and medium-level reading comprehension in 2017 and 2018 respectively, visual reasoning2 in 2020 and natural language inference3 in 2021 (Maslej et al., 2024[2]). Advancements in Natural Language Processing (NLP), and in particular in Large Language Models (LLM), enable applications like Generative AI to perform a wide range of language and cognitive tasks, often at a level comparable to humans and much faster. Generative AI refers to AI systems capable of creating new content based on patterns learned from existing data. For instance, ChatGPT, Gemini, or HyperCLOVA X in Korea, can write poems, computer code, and essays, compose music, and explain complex scientific ideas to a broader audience. When evaluated against answers given by experts on different questions, ChatGPT performance has been assessed as good as that of a team of experts (Guo et al., 2023[3]).
AI can now answer around 80% of the literacy and two‑thirds of the numeracy questions included in the OECD Survey of Adult Skills of the Programme for International Assessment of Adult Competencies (PIAAC). Comparing these results to those of the adults performing the tests highlights the potential for AI to outperform large portions of the adult population in reading and mathematics. Experts predict that increasing investments in AI research and development, in particular in NLP, will lead to further significant advancements of AI in both reading and mathematics over the coming years (OECD, 2023[4]).
Important progress has also been made in AI’s ability to replicate psychomotor abilities, specifically: the ability to work in cramped workspace, finger dexterity and manual dexterity. Finger dexterity refers to the ability to make precisely co‑ordinated movements of the fingers to grasp, manipulate, or assemble very small objects. Manual dexterity, by contrast, is the ability to quickly move the hand, the hand together with the arm, or the two hands to grasp, manipulate, or assemble objects. Older technologies, such as robots, are being improved through the integration of AI (Lassébie and Quintini, 2022[5]).
AI has made the least progress in physical abilities such as static, dynamic, and trunk strength.4 These are tasks more common in non-cognitive, non-routine occupations such as dancers, athletes, bricklayers, and farm workers.5 There are also other skills and abilities that humans still have a comparative advantage in, such as negotiation, social perceptiveness, assisting and caring for others, originality, and persuasion. Bringing people together and reconciling different views, understanding why people react a certain way, or providing emotional support, all remain complicated tasks for machines to perform (Georgieff and Hyee, 2021[6]; Lassébie and Quintini, 2022[5]). As of 2023, AI fails to exceed human ability also in some more complex cognitive tasks, such as visual commonsense reasoning6 and advanced level mathematical problem solving (Maslej et al., 2024[2]).
Measures of AI exposure, such as the one constructed by Felten, Raj and Seamans (2021[7]), evaluate the overlap between the abilities required in an occupation and the technical capabilities of AI. The occupations most exposed to AI are white‑collar occupations, which are most likely to involve non-routine cognitive tasks requiring formal training and/or tertiary education, such as IT professionals, business professionals, managers, and science and engineering professionals. Occupations requiring manual skills and strength, such as cleaners, agricultural forestry and fishery labourers, food preparation assistants and labourers, are the least exposed to AI (Figure 2.1) (Lane, 2024[8]; Georgieff and Hyee, 2021[6]). Focusing on generative AI, Eloundou et al. (2023[9]) observe that most occupations exhibit some degree of exposure to LLMs. Occupations with higher wages and information processing industries exhibit high exposure, while manufacturing, agriculture, and mining industries demonstrate lower exposure (Eloundou et al., 2023[9]). Felten, Raj and Seamans (2023[10]) also find that occupations with higher wages are more likely to be exposed to rapid advances in language modelling, and that education and legal service sectors exhibit higher exposure. Box 2.1 presents data on AI adoption based on survey results.
Average AI exposure by occupations, 2022
Source: Average AI exposure scores retrieved from Lane (2024[8]), “Who will be the workers most affected by AI?: A closer look at the impact of AI on women, low-skilled workers and other groups”, https://doi.org/10.1787/14dc6f89-en.
AI is increasingly recognised as a transformative technology with the potential to significantly impact workplaces. As a result, there is growing interest in understanding how widely these technologies are adopted by companies. According to Information and Communication Technology (ICT) surveys conducted by National Statistical Offices in 2024, the average AI adoption rate across OECD countries is 14%. Adoption rates vary by firm size. On average across OECD countries, 40% of large firms use AI, compared to 20% of medium-sized firms and just 12% of small firms (OECD, 2025[11]). For some countries, the latest available data predate the release of ChatGPT and other forms of generative AI. If more recent data were available for these countries, the average OECD AI adoption rate would likely be higher.
While European surveys benefit from standardised questions that facilitate cross-country comparisons, surveys from other regions often differ in design and definitions, posing challenges to comparability. New data on AI adoption in SMEs (small and medium-sized enterprises) has emerged from an OECD survey conducted between October and December 2024 examining the impact of generative AI on SMEs’ labour and skill needs (OECD, 2025[12]). The results highlight significant cross-country differences in AI adoption by SMEs, ranging from 27% in Japan to 51% in Germany. Korea is on the lower end of the scale, with 31% of SMEs saying they’ve adopted AI (Figure 1.2).
So far, in OECD countries, AI seems to have had no significant negative impact on overall employment
High exposure to AI does not necessarily imply workers in these occupations will be displaced. Theoretically, there are various channels through which the introduction of AI in the workplace could impact labour demand. Firstly, AI can substitute workers by automating tasks previously performed by human labour (displacement effect). Secondly, as some tasks are automated and AI can complement workers helping them perform tasks more efficiently, productivity increases and costs are reduced. This leads to lower quality-adjusted prices, potentially increasing product/service demand and, consequently, the demand for workers essential in the production process (productivity effect). Lastly, AI can create new tasks and jobs, particularly in AI development and maintenance (reinstatement effect). Therefore, the overall effect of AI on labour demand is theoretically ambiguous and depends on which effects dominate (Acemoglu and Restrepo, 2019[13]). To understand the impact of AI on aggregate employment empirical research is needed.
So far, across OECD countries, there is little evidence of negative aggregate employment outcomes due to AI. Instead, there appears to be a slight positive relationship between AI exposure and employment growth, suggesting that AI may be creating more jobs than it is destroying. At the same time, specific AI technologies could have different, and in some cases negative, impacts. What most studies highlight, is that while more jobs may be impacted by AI, very few are at risk of disappearing entirely. Most occupations involve a combination of skills and abilities that can and cannot be automated. Even highly impacted occupations are unlikely to be fully replaced by automation. Instead, work may need to be organised differently, and workers in these roles may require retraining as technology takes over certain tasks (Lassébie and Quintini, 2022[5]).
Case studies carried out by the OECD in the finance and manufacturing sectors of 8 OECD countries7 in 2022 showed that for 23% of the firms interviewed, AI technologies reduced the number of jobs in the most affected occupations. However, most firms managed these reductions by reallocating workers within the company or through attrition, keeping employees until they either left voluntarily or retired. In addition, firms often opted to slow hiring instead of implementing job cuts, using this approach as a safeguard against the potential failure or underperformance of AI solutions (Milanez, 2023[14]). This is consistent with the finding by Acemoglu et al. (2022[15]) that firms more exposed to AI reduce their overall hiring. Only a handful of studies, exploiting variation in AI adoption across US commuting zones, have found a negative effect of AI exposure on employment overall (Huang, 2024[16]; Bonfiglioli et al., 2025[17]).
The majority (77%) of the firms participating in the aforementioned case studies reported no impact on the quantity of jobs for workers most affected by AI technologies. Half of these firms implemented AI technologies to boost production volumes or improve product or service quality, rather than to reduce labour costs. For the other half, the implementation of AI led to the reorganisation of jobs, with workers displaced from certain tasks reassigned to other existing or new tasks. In some cases, AI technologies automated tasks that constituted only a minor share of workers’ jobs, thus not leading to displacement. In other cases, job reorganisation affected more substantial shares of workers’ tasks. However, these jobs were not eliminated, and the automation of certain tasks allowed workers to focus on more complex tasks that could not yet be automated (Milanez, 2023[14]). Additionally, 83% of SMEs report that the use of generative AI has had no effect on the overall number of staff they need (OECD, 2025[12]). Similarly, several studies do not find a significant relationship between AI exposure and aggregate employment (Felten, Raj and Seamans, 2019[18]; Georgieff and Hyee, 2021[6]; Acemoglu et al., 2022[15]). However, it is possible that significant impacts on aggregate economic data only become detectable once the technology is widely adopted and the necessary complementary processes and assets are developed (Brynjolfsson, Rock and Syverson, 2017[19]; Acemoglu et al., 2022[15]; Lane, 2024[8]).
Studies by Albanesi et al. (2023[20]) and Lane (2024[8]) have found a small, positive and statistically significant effect of AI exposure on aggregate employment, although direct causality is difficult to prove. The positive association between AI exposure and employment could be due to a productivity effect, or because AI creates new jobs directly. Green and Lamby (2023[21]) find that employment growth for the AI workforce, defined as workers with the skills necessary to develop and maintain AI systems, is strong. On average employment growth was 63% for the AI workforce between 2017 and 2019–although this workforce is still relatively small, representing less than 0.3% of workers overall. Acemoglu et al. (2022[15]) also find a rapid take‑off of AI vacancy postings starting in 2010 and accelerating around 2015‑2016. Moreover, in 30% of the OECD case studies, interviewees noted that employment was increasing in occupations related to the development and maintenance of AI (Milanez, 2023[14]).
AI could help mitigate labour shortages
Labour shortages are becoming a critical concern across many OECD countries. Labour market tightness, measured as the number of vacancies per unemployed person, has eased in the last quarter of 2023 but continues to exceed pre‑COVID‑19 levels in many countries (OECD, 2024[22]). Population ageing is a significant factor contributing to this challenge. As the workforce shrinks and demand for services like healthcare grows, innovative solutions are needed to avoid significant skills and labour shortages.
AI could help address these challenges by automating tasks and by enhancing worker productivity, enabling a more efficient use of resources, and making organisations better equipped to manage with a reduced workforce. AI could also support healthcare professionals by, for example, serving as a documentation assistant reducing the time spent on administrative tasks, or by assisting radiologists in scanning medical images, freeing up time for doctors to spend on care (Anderson and Sutherland, 2024[23]).
Furthermore, AI could help extend working lives and increase the labour market participation of the elderly. Many physically demanding jobs can lead to muscular-skeletal problems, but AI could be used to protect workers from injury, enabling them to remain employed longer. For example, the company German Autowerks invested in AI to analyse videos of mechanics at work, identifying pressure points and potential problem areas on the body. This information was then used to select specific exoskeletons which make heavy tasks much easier to perform. The company opted not to automate these tasks, believing that humans are more flexible and adaptable to changing job requirements. Instead, they focussed on using AI to assist workers, helping them work for longer (Machin, 2024[24]).
Despite the potential of AI to address challenges associated with labour shortages, it can only be part of a wider package of solutions to tackle these issues. Even if AI, particularly since the advent of Large Language Models, could be applied to a substantial share of tasks done by workers (Eloundou et al., 2023[9]), there are still tasks AI cannot do (or that society would not find acceptable for AI to do) and it cannot therefore fully replace workers (Lassébie and Quintini, 2022[5]). In addition, while several experimental studies show that AI could significantly enhance worker productivity in certain tasks (see section on Equalisation of performance within occupations below), the extent of this impact on aggregate productivity remains a topic of debate. This uncertainty is reflected in the ‘’productivity paradox’’, which refers to the lag in productivity growth over the past decade despite advancements in AI and other technologies. One possible explanation is that the aggregate productivity gains from AI might be modest (Acemoglu, 2024[25]). Alternatively, delays in AI implementation and organisational restructuring could mean that substantial economic gains from AI may take years or even decades to materialise (Lane and Saint-Martin, 2021[26]).
In OECD countries, AI has increased the need for new skills, including specialised AI and analytical skills
The integration of AI into the workforce has expanded the demand for specialised AI skills needed for the development and maintenance of AI systems. However, these positions still only represent a small fraction of total employment (Green and Lamby, 2023[21]). Most workers will have to interact with AI applications which often feature user-friendly interfaces, requiring only basic digital skills. The demand for skills complementary to AI – such as cognitive, management, social, and digital skills – appears to be increasing overall, while that for routine skills might decrease. Nonetheless, research also suggests these increases might not be related to AI per se, and AI exposure might be associated with a fall in demand for some of those skills.
The demand for AI skills in the labour market is increasing. Using data on skill requirements in online vacancies, Alekseeva et al. (2021[27]) show that the demand for these skills quadrupled over the period 2010 to 2019. The skills most demanded in AI vacancies are machine learning, natural language processing, deep learning, image processing, programming languages like Python, and big data management (Alekseeva et al., 2021[27]; Manca, 2023[28]; Squicciarini and Nachtigall, 2021[29]). These skills will be needed not only to design algorithms, but also to explain their functioning to non-technical professionals, and to monitor outcomes to make sure that AI systems are operating as intended, detecting mistakes and potential biases, and addressing any unintended consequences (Wilson, Daugherty and Morini-Bianzino, 2017[30]). Job postings that require specialised AI skills also tend to ask for high-level cognitive skills such as creative problem solving, social skills and management skills (project and people management), suggesting that these skills are complementary to AI. Conversely, these jobs typically do not require routine skills, like general administrative and clerical skills. As a result, an increase in AI-related employment is likely to drive demand for high-level cognitive skills while decreasing demand for routine skills (Alekseeva et al., 2021[27]; Manca, 2023[28]).
Nonetheless, most workers who will interact with AI may not need AI-specific skills or a deep understanding of AI systems. A survey of AI start-ups found that only 10% required users of their AI products to have expert coding or data skills, while 59% required only general computer familiarity, and the remainder required no specialised skills at all (Bessen et al., 2023[31]).
In many cases, AI adoption has not yet significantly changed skill requirements within firms. In 2022, 57% and 48% of firms that had adopted AI in finance and manufacturing, respectively, reported no change in skill needs (Lane, Williams and Broecke, 2023[32]). Similarly, in case studies of firms having implemented AI in finance and manufacturing, 60% of firms said that AI adoption had not yet modified skill requirements (Milanez, 2023[14]). This could be partly because AI adoption at the time of those studies was still relatively low and many firms were only experimenting with the technology, but also because interacting with AI applications often requires only basic digital skills, such as the ability to use a computer or smartphone, relying on existing skills.
That being said, 40% of the firms interviewed as part of the above‑mentioned case studies reported a need for new skills, including specialised AI skills and analytical skills. As simple tasks become automated, the proportion of complex tasks performed by workers rises, necessitating specialised knowledge and advanced analytical skills, such as the ability to comprehend and apply new ideas (Milanez, 2023[14]). Managers using algorithmic management software report that the use of such tools mostly increases their need for the ability to use or interpret data, and for digital skills (Milanez, Lemmens and Ruggiu, 2025[33]). Employers also say that, while AI has increased the importance of specialised AI skills, it has increased the importance of human skills, such as creativity and communication, even more, as well as the need for highly educated workers more generally (Lane, Williams and Broecke, 2023[32]). Green (2024[34]) shows that occupations with high AI exposure predominantly demand management, business processes, social and digital skills, with the largest increase in demand for skills related to collaboration, originality, and basic office tools.
At the same time, there is tentative evidence of a relative decline in the demand for management, business process, cognitive, digital, emotional and communication skills in workplaces that are highly exposed to AI (Green, 2024[34]). These skills are the skills of white‑collar support occupations: finance, human resources, legal, communications, administrative assistants and project managers. These effects are modest and should be viewed as relative changes amidst an overall increase in demand for most of these skills in the aggregate (Green, 2024[34]). In addition, however, managers using algorithmic management tools are more likely to report decreases in human interactions than increases. In the European countries surveyed,8 managers were also more likely to believe that the use of such tools was decreasing managers’ need for empathy rather than increasing it (Milanez, Lemmens and Ruggiu, 2025[33]).
These findings point to an additional concern, which is the potential deskilling of the workforce as a result of AI adoption. The case studies carried out by the OECD in the manufacturing and finances sectors of eight OECD countries documented some instances of deskilling, where the machine performed the skilled tasks, and the worker was only required to operate a very intuitive system, with no judgment involved (Milanez, 2023[14]).
Most OECD countries will need to ramp up training provision to address AI-induced skills demand
The changing skills needs resulting from the adoption of AI in the workplace call for new training opportunities. While initial education plays a crucial role in equipping workers with the skills to work with AI, upskilling and reskilling the existing workforce will be equally important to help individuals adapt and prepare for the transition (OECD, 2024[35]). Older adults and lower skilled workers in particular will need to acquire basic digital skills essential for interacting with AI technologies. Meanwhile, managers and business leaders require training to efficiently organise the integration of AI into their operations.
More than half of workers who use AI in the manufacturing and finance sectors said that their company had either provided or funded training so that they could work with AI. Yet, more training would help address existing barriers to AI adoption, considering that around 40% of employers in those sectors declared that the lack of relevant skills was a barrier to AI adoption (Lane, Williams and Broecke, 2023[32]). When the AI technology is simple to use, training can be brief and take the form of webinars, presentations, or workshops (Milanez, 2023[14]). In a small number of case studies, large firms operated more ambitious training programmes to help employees transition to other occupations. Some large companies try to grow AI talent in-house instead of seeking those employees on the external labour market. However, several firms call for more government funding for AI education and training, recognising that these specialised AI skills should also be developed in initial education (Milanez, 2023[14]).
Initial education plays a crucial role in acquiring the skills necessary to develop and maintain AI systems, with two‑thirds of the AI workforce holding a tertiary degree. The share of AI workers who report having participated in some sort of training in the last four weeks is similar to that of the entire population with a tertiary degree (16% and 18%, respectively). Most of the training that the AI workforce undertakes is non-technical in nature. Nonetheless, workers who have skills closely related to AI skills may acquire more explicit AI skills simply by being part of a research team or the AI development process within their firms (Green and Lamby, 2023[21]).
Lower-skilled and older workers are less likely to possess the basic digital skills required in a workplace transformed by the adoption of AI. Based on the Survey of Adult Skills, that tests adults in basic information processing skills, around one in four adults (aged 16‑65) have no or only limited experience with computers or lack confidence in their ability to use them. Additionally, nearly half of all adults can only use familiar applications to solve problems involving few steps and explicit criteria, such as sorting emails into pre‑existing folders. Among low-educated adults, 41% lack basic proficiency in using information and communications technology (ICT) to even take the survey’s test, and those who can undertake the test perform poorly. The percentage of adults without basic ICT skills decreases to 15% for those with upper secondary education and 4% for those with tertiary education. Compared to younger adults (aged 25‑34) older adults (aged 55‑65) are significantly more likely to have no computer experience and lower scores. Among young adults, 8% have no computer experience and 43% perform well. In contrast, 34% of older adults have no computer experience, and only 10.3% perform well (OECD, 2019[36]).
Moreover, older and lower-skilled adults, as well as low-wage workers, are less likely to take part in adult learning in every single country participating in the Survey of Adult Skills. Considering that around half of all adults neither participate nor want to participate in adult learning, it will be crucial to find effective ways to address barriers and motivation to training participation (OECD, 2019[37]). For example, the provision of more flexible learning options (e.g. part-time study or online delivery) would allow learners to better balance training alongside work or other commitments (OECD, 2024[35]).
Managers also need training to effectively organise the integration of AI into their operations. They need to understand AI systems to assess where and how innovation can be utilised within the company, identify the benefits and risks of AI, and determine the best ways to integrate AI systems into existing processes. Managers would have to decide which tasks are better performed by AI systems and which by humans, recognising the strengths and weaknesses of each (OECD, 2023[38]). A German insurance provider reported that managers planning AI projects are expected to have a minimum knowledge of how the technology works (Milanez, 2023[14]). In the OECD survey on algorithmic management in the workplace (2025[33]), 75% of managers report their firms offer training on how to use the software. Training for managers could enhance their proficiency with specific tools, deepen their understanding of the data used by these tools, and ensure their skills keep pace with the growing demand for analytical capabilities. Additionally, it could help them use software in a trustworthy manner (Milanez, Lemmens and Ruggiu, 2025[33]).
Encouragingly, several OECD countries have developed dedicated AI training strategies (OECD, 2024[35]). Many have introduced incentives to support employers in providing training for their employees. Fourteen governments have invested in publicly funded AI training programmes, with nine focussed on developing AI professionals and seven aimed at enhancing AI literacy for the general public. More broadly, publicly funded digital skills training, without an explicit focus on AI, is more common (OECD, 2024[35]).
The adoption and use of AI in the workplace in Korea is lower than in other countries
The adoption of AI in Korea appears low compared to other OECD countries (9.9% in 2024 from Han (2023[39]), 6.35% in 2023 from 2024 Survey of Business Activities (KOSIS, 2025[40]), 30.3% in 2023 from the 2024 Enterprise Informatization Statistics (NIA, 2025[41]) and see Figure 1.2) and firms appear to take a cautious stance towards AI, as shown by their investments in AI (Box 2.2).
The majority of firms that participated in the Survey on AI Utilisation and Labour Market Changes conducted by the Korea Labor Institute invest 5% or less of their sales revenue in AI development (Figure 2.2, Panel A). Among the companies that indicated they would increase AI investment in the future, a majority (55.3%) reported that their investment would be limited to 5% or less of their total sales revenue, suggesting a cautious stance on AI development expenditures in Korea (Figure 2.2, Panel B). This stance toward AI is due to concerns about the maturity of the technology and uncertain returns. Many firms prefer a gradual and incremental approach.
Percentage of firms reporting investment of 5% or less (of sales revenue), 6‑10%, 11‑20% and 21% or more
Note: The survey targeted firms that use AI, focussing on industries classified under the Korean Standard Industrial Classification (KSIC), specifically: Manufacturing, Information and communication, Professional scientific and technical service, Healthcare. Only firms that utilise AI and have 10 or more employees were included in the survey. HR managers and AI developers provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The survey covered a population of 9 625 establishments, including 3 292 in manufacturing, 3 118 in information and communication, 1 788 in professional and scientific services, and 790 in healthcare. The sample was drawn using a random sampling method, with a target sample size of 200. Ultimately, the study achieved valid responses from 145 firms, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes conducted by the Korea Labor Institute (2024).
Even if AI is adopted by organisations in Korea, that does not mean that employees use it frequently. 34.5% of workers who use AI, use it once or twice a week, while 32.2% used it once or twice a day (Figure 2.3). According to the Survey on AI Utilisation and Labour Market Changes conducted by the Korea Labor Institute as well as qualitative evidence, employees within the same team or department differ significantly in how frequently and intensively they use AI. AI users typically begin experimenting with the technology out of curiosity. As they gain experience, they become more comfortable and gradually expand their use of AI across various tasks. Users emphasise that AI reduces the time and effort needed to visualise design concepts, allowing for greater efficiency. In contrast, non-users tend to avoid engaging with AI altogether. They cite two main reasons: first, they see no meaningful application of AI in their specific tasks; second, they believe their current manual methods are faster and more accurate. This variation in AI adoption, even among employees performing identical work, highlights the importance of understanding individual-level factors that influence technology acceptance in the workplace.Box 2.3 provides a concrete example of this in a publishing firm in Korea.
Percentage of employees reporting using AI 1‑2 times a week, 1‑2 times every day, 1‑2 times a day, countless times
Note: The survey targeted individual employees who use AI in firms operating within four industries classified under the Korean Standard Industrial Classification (KSIC): Manufacturing, Information and Communication, Professional, Scientific and Technical Services, and Healthcare. Employees using AI provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The sample was drawn using a random sampling method, with a target sample size of 600. Ultimately, the study achieved valid responses from 426 employees, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes conducted by the Korea Labor Institute (2024).
A Korean publishing firm employing around 100 people promotes the adoption of AI in the workplace but does not require employees to use it. AI is applied across multiple departments. A notable example is a team responsible for designing book covers, where designers use image‑generation AI tools. Even within this team, the level of AI use varies. The team is composed of five designers; two actively use AI, whereas the other three do not. Those who use AI often begin out of curiosity. As they gain experience, they become more comfortable with the tools and gradually expand their use across different tasks. Designers who do not use AI rarely attempt to engage with the technology. AI users emphasise the reduced time and effort required to visualise design concepts. They report that AI enables them to generate images more quickly and efficiently. In contrast, non-users argue that their manual design methods offer superior quality and greater control.
While total employment continues to grow in Korea, some forms of more traditional AI appear associated with less growth in full-time, permanent employment
So far, in Korea, most research either finds no or a small negative impact of AI on employment overall, and the impact depends on both the type of technology and the industry. Chang et al. (2024[42]) concludes that AI-adopting firms experience productivity gains without reducing employment while Han (2023[39]) finds employment increases in high-skilled jobs and losses in low-skilled jobs. New analysis carried out for this report shows that more “traditional” AI appears to be associated with lower growth in full-time, permanent jobs, concentrated in the manufacturing sector, while there is no such association with generative AI (Box 2.4). In particular, a 10‑percentile point increase in exposure to “traditional” AI (Webb, 2020[43]) appears to be associated with 5.7% lower growth in full-time, permanent employment. This needs to be interpreted in a context of 2.4% employment growth in full-time, permanent jobs overall during the same period, as well as 5.8% growth in total employment (including non-standard forms of work).
This box presents new analysis of the impact of AI on employment growth at the occupational level in Korea, spanning the period from 2018 to 2023. The analysis draws on three primary data sources: employment insurance data, exposure indices from the studies of Webb (2020[43]) and Felten et al. (2023[10]), and the routinisation index extracted from the 2020 Korea Dictionary of Occupations, published by the Korea Employment Information Service (2020[44]).
Employment insurance data contain detailed individual-level information, including worker demographics (e.g. age, gender, career history), occupation, and wages, as well as information regarding the region, industry, and size of the establishments where individuals are employed. Notably, the employment insurance data include annual information on individual labour market transitions and the diverse characteristics of workplaces. However, as this dataset is specifically used for the administration of employment insurance, it only covers individuals and workplaces enrolled in the scheme. In Korea, employment insurance typically applies only to permanent employees, resulting in a potential bias, as the data predominantly reflects regular, full-time workers (Kwon, 2022[45]). According to the Economically Active Population Survey, approximately 25% of employed individuals in Korea were temporary workers as of 2023.
Two distinct AI exposure indices were employed: one developed by Webb (2020[43]) and the other by Felten et al. (2023[10]). Webb’s (2020[43]) index measures AI exposure by extracting job-related information from AI patents, representing exposure to more “traditional” AI technologies. In contrast, Felten et al. (2023[10]) assess exposure based on the degree to which occupations are affected by AI language modelling capabilities, providing a measure of exposure to generative AI technologies.1 To account for other automation technologies that influence labour demand and supply within occupations, it was necessary to control for their effects, distinct from those of AI. Accordingly, the analysis incorporates software exposure and robot exposure at the occupational level, as calculated by Webb (2020[43]).
While existing studies on AI and automation technologies typically focus on the US labour market, variations in occupational structures between countries such as Korea, the US, and EU and its high-income constituent members, can arise due to diverse factors like industrial structure. To control for these potential differences and to reflect the exposure to pre‑AI automation technologies within Korea’s occupational classification system, a routinisation index was used. This index, first introduced by Kim, Koh and Cho (2014[46]), categorises occupations according to three primary attributes – data, people, and “objects” (i.e. the various tangible assets employed during the execution of job duties)–based on the 2020 Korea Dictionary of Occupations.2 Each occupation is assigned a score between 0 and 2 for each characteristic, with the highest score determining the final routinisation index. A higher value indicates lower exposure to routinisation, while a lower value suggests greater exposure.
For the analysis, raw values of the AI exposure indices, as well as the routinisation index, were converted into percentiles. This approach mitigates potential distortions caused by the clustering of values within the raw indices, facilitating a clearer understanding of the differences in occupational exposure. Consequently, the results focus on changes in the dependent variables in response to shifts in percentiles, rather than the raw values of the indices.
Table 2.1 presents the findings of the analysis, which explores the impact of the two AI exposure indices on the growth in full-time, permanent employment in Korea’s labour market over the 2018‑2023 period. The year 2018 was chosen as the baseline, as AI adoption in Korea was minimal prior to this period.
Model (1) controls for various occupational and industry-level characteristics, such as gender ratio, average monthly wages, and age distribution. Model (2) incorporates the software exposure index, while Model (3) includes the robot exposure index. Model (4) further controls for the routinisation index, and Model (5) accounts for trends in the dependent variable prior to the analysis period (2015‑2018). Unless otherwise noted, the results discussed herein refer to the outcomes from Model (5).
Note: AIOE stands for AI Occupational Exposure Standard errors are shown in parentheses, and statistical significance is indicated as follows: *** 1%, ** 5%, * 10%.
Source: Employment insurance data and information provided by Webb (2020[43]), “The Impact of Artificial Intelligence on the Labor Market”, https://www.michaelwebb.co/webb_ai.pdf; and Felten et al. (2023[10]), “How will Language Modelers like ChatGPT Affect Occupations and Industries?”, https://doi.org/10.48550/arXiv.2303.01157 compiled by the authors.
The results show that a 10 percentile point increase in exposure to traditional AI is associated with a 5.7% decline in full-time, permanent employment growth, whereas no statistically significant relationship exists for generative AI. Given that there has been a 2.4% increase in permanent employment over this period (2018 to 2023), these findings suggest that there has been employment growth at median levels of exposure to traditional AI, but that this positive effect reduces as exposure rises and that the displacement begins to outweigh the productivity gains at higher levels of exposure to traditional AI.
1. Although Felten et al. (2018[47]) also created an index for traditional AI through expert surveys, analyses using this index did not yield results significantly different from those obtained from Webb’s exposure index. Therefore, for the purpose of this study, only the Webb and Felten indices are employed to distinguish the impacts of traditional and generative AI technologies on the labour market.
2. The 2020 Korea Dictionary of Occupations breaks down each occupation into three characteristics: the use of data, interaction between people, and the use of things. In other words, “objects” collectively refers to the various types of tangible assets employed during the execution of job duties.
At the same time, a recent survey showed that 95.5% of firms in Korea reported no workforce changes at the department- or team-level following the adoption of AI. Case studies carried out in Korea suggest this may be because AI adoption still remains relatively low (see Box 2.5). Among the firms that reported changes in tasks as a result of AI adoption, only 8.4% reported full automation of all tasks in a job (Figure 2.4). In addition, 35.1% of firms indicated that AI had been adopted to perform entirely new tasks that had not previously existed within the scope of traditional roles, suggesting that AI adoption may not necessarily be geared towards replacing workers.
Percentage of firms reporting that AI creates entirely new tasks, partially replaces tasks, or fully replaces tasks in a job
Note: The survey targeted firms that use AI, focussing on industries classified under the Korean Standard Industrial Classification (KSIC), specifically: Manufacturing, Information and communication, Professional scientific and technical service, Healthcare. Only firms that utilise AI and have 10 or more employees were included in the survey. HR managers and AI developers provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The survey covered a population of 9 625 establishments, including 3 292 in manufacturing, 3 118 in information and communication, 1 788 in professional and scientific services, and 790 in healthcare. The sample was drawn using a random sampling method, with a target sample size of 200. Ultimately, the study achieved valid responses from 145 firms, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes conducted by the Korea Labor Institute (2024).
KLI carried out interviews with HR and technology managers from firms using AI within the industries covered by the Survey on AI Utilisation and Labour Market Changes, i.e.: manufacturing (2 interviewees), information and communications (5), healthcare (3), and professional, scientific, and technical services (5). The interviews were conducted over approximately one month, from 24 February 2025 to 26 March 2025, either in person or online, depending on the interviewees’ preferences. The interviews paint a nuanced picture of AI’s impact on automation. While job losses may happen in the future, automation is currently limited partly because of low adoption but also because AI tends to automate only parts of, and not the entire job.
Case study 1: Intellectual Property Dispute Resolution
One company in Korea has developed an internal AI tool for foreign language translation and utilises an internal version of ChatGPT. The AI tool assists with translating foreign language documents (with 90% accuracy), searching case law, and drafting legal documents. The implementation of AI has not led to any changes in the team’s composition. Employees perceive that AI can perform only about 2 out of 10 tasks in their overall workload. However, as technology advances, they anticipate that fewer team members will be needed for the same tasks in five to six years.
Case study 2: Radiology Technologist
In the field of pathology, AI adoption is more advanced than in other areas due to the vast amount of accumulated data, allowing for extensive AI training. As a result, AI is more actively utilised in pathology than in any other hospital department. However, one major limitation remains: since organ placement and size vary across patients, achieving consistently accurate results is still challenging. With further technological advancements, AI is expected to not only assist in imaging procedures but also take over diagnostic interpretations. Consequently, the demand for radiologic technologists may decline, as fewer professionals will be required to perform these tasks.
Case study 3: Manufacturing
Most of the AI models adopted in a manufacturing firm in Korea are used as references for the corresponding production workers. Workers can shut down and inspect the equipment by an alert of an AI to a potential equipment failure. Previously, workers would intuitively detect such failures based on noise, temperature, vibration, or other factors. Now, the AI senses the failure by combining relevant data. If the AI is accurate and reliable, stricter criteria can be set for the failure threshold. If not, looser criteria are used. The strictness of these criteria varies across AI models, but many models in this firm are said to have somewhat loose criteria. This suggests that most AI models are used as references rather than as a judge which completely replaces human labour. It is also worth noting that even in cases where AI replaces human labour, only part of the job (one or two tasks) is replaced, not the entire role. In this sense, AI and human labour work together, and the adoption of AI models rarely affects the overall employment size of the firm.
The limited impact of AI on human labour stems primarily from the lower accuracy and reliability of the AI models in this manufacturing firm. Therefore, this does not necessarily mean that AI will not have a negative impact on employment in the future. As AI technology advances, and as the quality and quantity of data improve, AI will become more accurate and reliable. Given the complexity of work processes, the active involvement of production workers as domain knowledge experts is crucial to improving the quality of AI learning outcomes. Data quality and data-handling capabilities remain essential in AI development, even with more advanced AI techniques. However, the involvement of workers in advancing AI further in this firm does not seem to be sufficient. To enhance worker co‑operation, management needs to ensure that AI adoption will not negatively affect employment. This is a contradiction, because AI utilisation is typically aimed at reducing labour costs.
In line with the above findings, a recent OECD survey (2025[12]) shows that most Korean SMEs report no effect of generative AI on the overall number of staff (88%), the number of highly skilled staff (85%), or the number of lower skilled staff (87%). Also in line with the findings reported above, among the SMEs that have experienced an impact, firms are more likely to report an increased need for highly skilled workers than a reduction, and more likely to report a reduced need for low-skilled staff than an increase (Figure 2.5) (OECD, 2025[12]).
Percentage of SMEs reporting generative AI has impacted the company staffing needs
Note: The total reflects the combined results from all countries participating in the survey (Austria, Canada, Germany, Ireland, Japan, Korea, the United Kingdom).
Source: OECD (2025[12]), Microdata from the OECD SME Survey on Generative AI.
AI is already changing the tasks that workers do in Korea
AI changes the tasks that workers do. In a survey of Korean firms that have adopted AI, 56.5% reported that AI replaced specific tasks within existing jobs (rather than automating the entire job) (see Figure 2.5 above). Moreover, 71% of Korean firms reported that AI substituted approximately 10% of an employee’s tasks, and 17.2% reported that AI had not replaced any of their tasks at all (Figure 2.6).
Percentage of firms reporting that AI replaces 0%, 10% or fewer, 11‑30% and 31‑50% of tasks
Note: The survey targeted firms that use AI, focussing on industries classified under the Korean Standard Industrial Classification (KSIC), specifically: Manufacturing, Information and communication, Professional scientific and technical service, Healthcare. Only firms that utilise AI and have 10 or more employees were included in the survey. HR managers and AI developers provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The survey covered a population of 9 625 establishments, including 3 292 in manufacturing, 3 118 in information and communication, 1 788 in professional and scientific services, and 790 in healthcare. The sample was drawn using a random sampling method, with a target sample size of 200. Ultimately, the study achieved valid responses from 145 firms, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes conducted by the Korea Labor Institute (2024).
Unlike previous automating technologies, AI can also automate non-routine, cognitive tasks. Case studies of lawyers, physicians and webtoon creators in Korea show that while AI contributes to enhanced efficiency in tasks such as document drafting, diagnostics, and visual rendering, it does not yet fully substitute for human involvement in strategic decision making, interpersonal communication, or creative processes (Box 2.6).
Between 24 February 2025 and 26 March 2025, KLI carried out interviews with HR and technology managers from firms using AI. These interviews showed how, regardless of the occupation where AI was used, the technology only automated some tasks while leaving other tasks untouched.
Case study 1: Lawyers
AI is being used to assist with legal document drafting and case law analysis. However, tasks such as legal consultation, courtroom representation, and negotiation still require human expertise. These complex, judgment-based activities remain beyond AI’s current capabilities.
Case study 2: Physicians
AI supports diagnostics, medical testing, and treatment planning in clinical settings. Yet, tasks involving direct patient interaction, complex diagnoses, and surgeries remain firmly in the domain of human professionals. The need for empathy, intuition, and clinical judgment limits AI’s role.
Case study 3: Webtoon creators
AI is used in tasks like storyboard generation, sketching, colouring, and adding special effects. Nonetheless, story development, character design, and audience interaction continue to depend on human creativity. These artistic and narrative aspects resist automation.
AI is increasing the demand for high-level and social skills in Korea
The changes in tasks brought about by AI will change the demand for skills too. A survey of firms in Korea reveals that 32.2% say the use of AI has resulted in an increase in the kinds of skills required to carry out current tasks, while 63.1% say there has been no change. Similarly, 38.3% of Korean firms said that AI had increased the level of skills required, compared to 55.9% of firms that said there was no change (Figure 2.7).
Percentage of firms reporting that (i) the kinds of skills and (ii) the level of skills required following AI adoption increase, decrease, do not change
Note: The survey targeted firms that use AI, focussing on industries classified under the Korean Standard Industrial Classification (KSIC), specifically: Manufacturing, Information and communication, Professional scientific and technical service, Healthcare. Only firms that utilise AI and have 10 or more employees were included in the survey. HR managers and AI developers provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The survey covered a population of 9 625 establishments, including 3 292 in manufacturing, 3 118 in information and communication, 1 788 in professional and scientific services, and 790 in healthcare. The sample was drawn using a random sampling method, with a target sample size of 200. Ultimately, the study achieved valid responses from 145 firms, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes Conducted by the Korea Labor Institute (2024).
These findings are consistent with what Korean SMEs say about skills needs changes due to generative AI. Overall, the demand for most skills is rising, with increases outweighing decreases. The importance of data analysis and interpretation skills increases the most, followed by programming and coding skills. The largest reported decline in importance is for critical thinking and problem-solving skills (Figure 2.8).
Percentage of SMEs reporting whether generative AI has made the skill more important or less important
Note: The total reflects the combined results from all countries participating in the survey (Austria, Canada, Germany, Ireland, Japan, Korea, the United Kingdom).
Source: OECD (2025[12]), Microdata from the OECD SME Survey on Generative AI.
In line with these findings, firms regardless of size in Korea are more likely to say that the adoption of AI results in an increase in the frequency of communication: among team members (18.4% v. 7.1%), with managers (19.1% v. 7.1%), as well as between teams (27% v. 7.8%) (Figure 2.9). These shifts may indicate a growing need for competencies typically classified as social skills, particularly in environments where collaborative decision making is essential. At the same time, while AI demonstrates strong performance in tasks such as data analysis and generating decision options, it remains limited in areas requiring emotional understanding, creative problem-solving, negotiation, and persuasion. These social skills are difficult to replicate through AI. As such, there is a growing argument that the integration of AI in the workplace increases the importance of social skills, such as communication and teamwork.
Percentage of firms reporting that the frequency of communication decreases, does not change, increases
Note: The survey targeted firms that use AI, focussing on industries classified under the Korean Standard Industrial Classification (KSIC), specifically: Manufacturing, Information and communication, Professional scientific and technical service, Healthcare. Only firms that utilise AI and have 10 or more employees were included in the survey. HR managers and AI developers provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The survey covered a population of 9 625 establishments, including 3 292 in manufacturing, 3 118 in information and communication, 1 788 in professional and scientific services, and 790 in healthcare. The sample was drawn using a random sampling method, with a target sample size of 200. Ultimately, the study achieved valid responses from 145 firms, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes Conducted by the Korea Labor Institute (2024).
By itself, AI will not solve labour shortages in Korea, but it can help mitigate them
Both the evidence on automation and changes in skills needs raise the question of whether AI can help firms compensate for worker shortage or lack of skills. Korea faces a rapidly ageing population. Although in 2022 only 17.5% of its population was 65 or older, this figure was 11.5% just a decade earlier and is steadily rising. This trend is driven by Korea’s fertility rate of 0.7 live births per woman in 2023, the lowest in the world (OECD, 2024[48]). Korea considers robots crucial for addressing the challenge of population ageing. The government has advocated for the establishment of a “K-robot economy”, encouraging the deployment of robots across various sectors including manufacturing, defence, aerospace, and services (Matsuura, 2024[49]). Many of these robots are likely to be integrated with AI systems to perform their tasks. SMEs in Korea experience significant labour shortages and generative AI might help in addressing these issues (OECD, 2025[12]).
In a survey of firms, 24.1% reported that AI played a role in mitigating labour shortages (Figure 2.10). These figures suggest that while AI by itself will not solve labour shortages, it can at least help mitigate the challenges. This finding is confirmed by a recent OECD survey of SMEs. 37% of Korean SMEs report experiencing a worker shortage in the last two years, and 27% of SMEs reporting shortages say that generative AI helps compensate for these shortages (Figure 2.11). Similarly, 24% of Korean SMEs report a lack of skills and experience among staff, and 47% of these SMEs say that generative AI helps to address this challenge.
Percentage of firms agreeing that AI solves workforce shortages
Note: The survey targeted firms that use AI, focussing on industries classified under the Korean Standard Industrial Classification (KSIC), specifically: Manufacturing, Information and communication, Professional scientific and technical service, Healthcare. Only firms that utilise AI and have 10 or more employees were included in the survey. HR managers and AI developers provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The survey covered a population of 9 625 establishments, including 3 292 in manufacturing, 3 118 in information and communication, 1 788 in professional and scientific services, and 790 in healthcare. The sample was drawn using a random sampling method, with a target sample size of 200. Ultimately, the study achieved valid responses from 145 firms, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes Conducted by the Korea Labor Institute (2024).
Note: The total reflects the combined results from all countries participating in the survey (Austria, Canada, Germany, Ireland, Japan, Korea, the United Kingdom).
Source: OECD (2025[12]), Microdata from the OECD SME Survey on Generative AI.
A key challenge in Korea is the brain drain of AI talent
Korea has experienced emigration of high-skilled workers since the 1950s. The main motives for migration have been social environment and political security, as well as a desire for higher living standards and better job opportunities (Jung, 2018[50]). However, as evidenced in the Artificial Intelligence Index Report 2025 (Maslej et al., 2025[51]), AI has introduced a new dimension to the brain drain. With long working hours and low wages compared to the OECD average, Korean AI experts have emigrated to the United States and high-income countries in Europe to seek better salaries and working conditions. Reduced costs of international mobility resulting from globalisation and advances in transport have exacerbated the outflow of high-skilled workers. The analysis of Korean employment insurance statistics showed that, in Korea, generative AI only led to increases in wages of STEM and high-paying occupations, while in the United States it led to increases in both wages and employment (Bonfiglioli et al., 2025[17]). This discrepancy likely reflects differences in the supply of AI experts: whereas the United States appears to have an adequate supply of AI professionals that meets increased demand, Korea has seen rising demand without a corresponding increase in labour supply, resulting primarily in wage increases for these workers.
In an effort to address this shortfall in skilled AI professionals, the Korean Government has implemented a range of AI-related educational initiatives and vocational training programmes. These programmes have been to some extent successful in supplying the market with workers possessing low to intermediate skill levels, but unsuccessful in providing high-skilled professionals, most of whom seek overseas job opportunities pursuing higher salaries and better working conditions. The primary concern is that this exodus undermines the nation’s AI competitiveness and threatens productivity. A shortage of AI experts slows down technological progress and productivity gains remain limited. Low productivity growth negatively affects corporate performance and firms’ ability to offer high salaries. This, in turn, exacerbates the shortage of skilled AI workers, creating a self-reinforcing cycle of brain drain and economic stagnation.
While the most effective way to retain AI talent would be to offer higher wages, firms in Korea have been reluctant to do so, relying on alternatives instead, such as non-pecuniary benefits. The Korean Government attempts to shore firms up by providing training for workers and education programmes for students. However, these efforts have not so far been successful in increasing the supply of high-skilled AI workers in the domestic labour market.
Participation in adult learning in Korea is the lowest across OECD countries
Participation in adult learning in Korea is the lowest among OECD countries: 13%, compared to an OECD average of 40%. In Korea, as in all other OECD countries, older and lower-skilled adults, as well as low-wage workers, are less likely to take part in adult learning (OECD, 2025[52]). At the same time, adult skills in Korea are low. 30% of adults (aged 25‑64) have no or limited experience with computers or lack confidence in their ability to use them. Additionally, 40% of Korean adults can only use familiar applications to solve problems involving few steps and explicit criteria, such as sorting emails into pre‑existing folders (Below Level 1 or Level 1) (Figure 2.12). Among low-educated adults in Korea, 76% lack basic proficiency in using information and communications technologies (ICT), and those undertaking the proficiency test perform poorly. The percentage of adults without basic ICT skills decreases to 7% for those with tertiary education (Figure 2.13, panel A). Compared to younger adults (aged 25‑34) older adults (aged 55‑65) are significantly more likely to have no computer experience and lower scores. Among young adults, 7% have no computer experience and 49% perform well. In contrast, 64% of older adults have no computer experience, and only 4% perform well (Figure 2.13, panel B) (OECD, 2019[36]).
Percentage of 16‑65 year‑olds scoring at each proficiency level
Note: The scale of problem solving in technology-rich environments is divided into four levels of proficiency (Levels 1 to 3 plus below Level 1).
Source: Survey of Adult Skills (PIAAC) (2012, 2015, 2018), (OECD, 2019[36]).
Note: The scale of problem solving in technology-rich environments is divided into four levels of proficiency (Levels 1 to 3 plus below Level 1).
Source: Survey of Adult Skills (PIAAC) (2012, 2015, 2018), OECD (2019[36]), Skills Matter: Additional Results from the Survey of Adult Skills, https://doi.org/10.1787/1f029d8f-en.
While firms in Korea do provide training to employees for working with AI, only 42% of those that have adopted AI have done so, and the share is higher in large firms than it is in small ones (Figure 2.14). Firms in the Manufacturing sector are also more likely to provide training than those in the Information and communication sector (32.6%).
Percentage of firms reporting providing training programmes to employees
Note: The survey targeted firms that use AI, focussing on industries classified under the Korean Standard Industrial Classification (KSIC), specifically: Manufacturing, Information and communication, Professional scientific and technical service, Healthcare. Only firms that utilise AI and have 10 or more employees were included in the survey. HR managers and AI developers provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The survey covered a population of 9 625 establishments, including 3 292 in manufacturing, 3 118 in information and communication, 1 788 in professional and scientific services, and 790 in healthcare. The sample was drawn using a random sampling method, with a target sample size of 200. Ultimately, the study achieved valid responses from 145 firms, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes conducted by the Korea Labor Institute (2024).
The quality of jobs can be assessed along a number of dimensions, including: the level and distribution of earnings; employment security; and the quality of the work environment, which includes aspects such as work intensity, performing physically demanding tasks, and worker autonomy (Cazes, Hijzen and Saint-Martin, 2015[53]). The introduction of AI in the workplace will affect many of these outcomes.
In OECD countries, the wage benefits of AI have been concentrated among high-income and highly skilled workers
At an aggregate level, there is no evidence to date in OECD countries of statistically significant relationship between AI exposure and wage growth. However, the effects vary across occupations and income groups, with workers employed in occupations with either high software prevalence or high incomes appearing to benefit from AI exposure in terms of higher wages.
The impact of AI on wages is theoretically ambiguous. Automation could reduce the need for workers in certain tasks, lowering demand for those occupations and putting downward pressure on wages. Conversely, increases in productivity, leading to cost reductions and lower prices, could boost demand for goods and services, increase labour demand and result in higher wages. Additionally, workers with specialised skills in AI development and maintenance would be in higher demand, which could lead to higher wages for this subset of workers.
At an aggregate level, empirical evidence shows little evidence of a relationship between wage growth and AI exposure (Acemoglu et al., 2022[15]; Albanesi et al., 2023[20]). Similarly, in case studies carried out by the OECD in the finance and manufacturing sectors of 8 OECD countries, 84% of interviewees reported that the wages of workers most affected by AI remained unchanged and 15% reported increases (Milanez, 2023[14]). One possible explanation for this lack of relationship is that AI advances may not lead to substantial increases in productivity. Acemoglu (2024[25]) predicts that total factor productivity (TFP) gains over the next decade will be modest, with an estimated annual increase of no more than 0.06%. However, such estimates are highly uncertain and depend on the underlying assumptions. A recent OECD study (Filippucci, Gal and Schief, 2024[54]) presents a more optimistic outlook, estimating aggregate gains for annual TFP growth between 0.25 and 0.6 p.p. Acemoglu (2024[25]) also anticipates a widening gap between capital and labour income, suggesting that the benefits of any productivity increases are likely to accrue more to capital owners than to workers. This implies that while AI may enhance productivity, the economic gains may not translate into significant wage increases for the labour force. These findings highlight the importance of social dialogue and collective bargaining in achieving a fair share of any productivity gains that might accrue (see section on social dialogue in Chapter 3).
Some studies find a positive relationship between AI and wage growth when focussing on specific groups of workers. Felten, Raj and Seamans (2019[18]) find that AI exposure correlates positively with wage growth, and that this result is largely driven by occupations requiring a high level of familiarity with software. The same study finds that the relationship between AI and wage growth is positive and statistically significant for high-income occupations but not for low- or middle‑income occupations. Fossen and Sorgner (2022[55]) find a positive relationship between AI exposure and wage growth, but which is statistically significant only for individuals with at least upper secondary education. They argue that education enhances the ability to learn new information allowing highly educated workers to adapt better to new technologies. These workers are also more likely to possess skills that cannot yet easily be replaced by digital technologies, such as creative and social intelligence, reasoning, and critical thinking skills.
With regards to the AI workforce, Alekseeva et al. (2021[27]) document a dramatic increase in the demand for AI skills over the period 2010‑2019 in the US economy across most industries and occupations (albeit from a very low baseline). This is associated with a wage premium of 11% for job postings that require AI skills within the same firm and of 5% within the same job title. However, according to Green and Lamby (2023[21]), wages grew cumulatively by 7% for the AI workforce compared to 9% overall over the period 2017 to 2019, suggesting that supply of AI workers may have been keeping up with the demand for them.
Evidence for OECD countries suggests that while AI can enhance job satisfaction and safety, it may also intensify work and increase stress
The automation of tedious and repetitive tasks by AI could improve job enjoyment and allow workers to focus on more complex and interesting tasks. AI could also improve physical safety by automating dangerous tasks and improving monitoring systems and safety procedure controls. However, there are concerns around increased work intensity and stress related to AI use in the workplace. Ultimately, the effect of AI on the work environment depends on how thoughtfully and strategically it is integrated into workplace practices. Firms that use AI to support and empower their employees are more likely to see positive outcomes than those that deploy AI in ways that increase pressure and reduce autonomy.
Workers who use AI are more than four times as likely to report that AI has improved working conditions than worsened them, across all indicators considered such as job satisfaction, physical health, mental health, and fairness in management. Most AI users reported that AI had assisted them with decision making, which workers appeared to appreciate (Lane, Williams and Broecke, 2023[32]). Similarly, managers using algorithmic management tools – technologies, including AI, to fully or partially automate tasks traditionally carried out by human managers – report improvements in the quality of their decision making and in their own job satisfaction. Beyond benefiting firms, improved decision making by managers may also benefit workers through greater consistency and objectivity of decisions (Milanez, Lemmens and Ruggiu, 2025[33]).
Similar findings about a positive impact of AI on working conditions emerged from more qualitative research, with case study interviewees often reporting fewer tedious and repetitive tasks as a result of AI adoption, allowing workers to spend more time on stimulating and interesting tasks, thus increasing job enjoyment. For example, chatbots can deal with a high percentage of basic and repetitive customer service queries, freeing up agents to deal with more complex questions (Moore, 2019[56]). Workers’ physical safety was improved following AI implementation as the automation of processes allowed dangerous machines to run within enclosures or behind barriers. Visual inspection tools used to ensure the quality of products reduced workers’ eye fatigue. Mental health improvements were noted as well. For example, in one company, an AI system monitored the stock of materials along an assembly line and automatically ordered replenishments when needed, relieving workers from this responsibility, and reducing stress. Another example is a predictive maintenance AI technology used in a drug manufacturing company to identify early signs of degradation in seals, which was welcomed for removing the pressure of making these decisions. The implementation of an AI-based tool in manufacturing helped solve maintenance issues along a production line, aiding in the discovery of root causes and suggesting solutions, thus allowing maintenance workers to solve problems quickly and reducing stress (Milanez, 2023[14]).
However, some reports indicated that AI could lead to a deterioration in job quality. Case study evidence showed some instances of increased work intensity due to higher performance targets or complexity induced by AI. An AI developer acknowledged this issue and chose to retain 10% of easy tasks, which could have been automated, to provide workers with necessary mental breaks (Milanez, 2023[14]). Survey evidence shows that three‑quarters of AI users said that AI had increased the pace at which they perform their tasks. While this could be related to increased worker productivity, it might also reflect increased work intensity. In finance 49% of workers and in manufacturing 39% reported that their company’s AI systems collected data on them as individuals or on their performance at work. Of these employees, 62% in finance and 56% in manufacturing experienced increased pressure to perform due to this data collection (Lane, Williams and Broecke, 2023[32]).
Some of the impact of AI on working conditions may be linked to the increased use of algorithmic management in the workplace. These tools can, for instance, be used to allocate work schedules and activities, or monitor and evaluate workers. Despite the potential benefits discussed above, some concerns remain. Over a quarter (27%) of managers expressed concerns about inadequate protection of workers’ physical and mental health (Milanez, Lemmens and Ruggiu, 2025[33]). When algorithms assign tasks and set deadlines, workers may feel pressured to accelerate their pace of work. This pressure can lead to stress and anxiety and, in the most extreme case, burnout or accidents in order to meet the deadlines set (Todolí-Signes, 2021[57]). Algorithmic management can also reduce workers’ autonomy and sense of control over their work. For example, warehouse workers are sometimes equipped with wearable devices that dictate what to collect, where to find it, and how long they have to retrieve each item, but also guide their movements through vibrations to be more efficient. If it takes longer than suggested, the worker may receive a warning notice. Depriving workers of the ability to make even minor decisions, or how to move their own limbs, reduces autonomy and can remove an important sense of dignity and humanity from work (Moore, 2018[58]; Brione, 2020[59]).
Ultimately, AI technologies can lead to different outcomes depending on the organisational context in which they are implemented. In hierarchical organisations, AI adoption can create an “algorithmic cage” limiting worker autonomy and increasing resistance to its use. In contrast, organisations that allow room for professional judgment can foster an “algorithmic colleague”, where AI is used as a supportive tool to enhance human decision making (Meijer, Lorenz and Wessels, 2021[60]).
In Korea, the wage benefits of AI have been concentrated on the occupations most exposed to generative AI
While traditional AI has had no significant impact on the wages of full-time, permanent employees so far, new analysis of employment insurance data suggests that generative AI may be associated with higher wage growth in occupations with the highest exposure (Box 2.7). There are two potential explanations for this result. One is that generative AI, as observed by Moon et al. (2023[61]), increases productivity, leading to higher wages. The other is that while labour demand for generative AI skills has increased, the novelty of the technology has led to an inelastic labour supply. This could result in an increase in wages without a corresponding change in the equilibrium quantity of labour, meaning employment remains stable while wages rise. While only market-traded labour and wage data are observed, and excess labour demand or supply are not directly captured, it is highly probable that labour demand for generative AI in the labour market is increasing.
This box extends the analysis of Box 2.4 to look at the impact of AI on wage growth among full-time, permanent employees. The results in Table 2.2 show that there is no statistically significant relationship between traditional AI (Webb index) and wage growth, while there is a positive association between generative AI and wage growth (Felten index).
Note: Standard errors are shown in parentheses, and statistical significance is indicated as follows: *** 1%, ** 5%, * 10%.
Source: Employment insurance data and information provided by Webb (2020[43]), “The Impact of Artificial Intelligence on the Labor Market”, https://www.michaelwebb.co/webb_ai.pdf; and Felten et al. (2023[10]), “How will Language Modelers like ChatGPT Affect Occupations and Industries?”, https://doi.org/10.48550/arXiv.2303.01157 compiled by the authors.
The impact of AI on other aspects of job quality in Korea is mixed
A survey conducted by the Korea Labor Institute in 2024 shows that AI appears to improve job performance and productivity, as well as job satisfaction. Firms and workers are equally positive about the impact of AI on job performance and productivity (Figure 2.15). However, there is a gap between the perceptions of firms and employees when it comes to job satisfaction. Firms gave an average score of 2.88, whereas employees reported a lower score of 2.67. Compared to firms, workers perceive less improvement in job satisfaction resulting from the use of AI.
Average score reported by firms and employees on how AI has influenced work outcomes
Note: Workers who use AI were asked to rate, on a 4‑point scale, the extent to which they experienced increases in job satisfaction, work accuracy, work convenience, task performance, and productivity due to AI use. The response options were: 1 (Not at all), 2 (Not really), 3 (Somewhat), and 4 (Very much so). The survey targeted firms and employees that use AI, focussing on industries classified under the Korean Standard Industrial Classification (KSIC), specifically: Manufacturing, Information and communication, Professional scientific and technical service, Healthcare. Only firms that utilise AI and have 10 or more employees were included in the survey. HR managers, AI developers and individual employees provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The sample was drawn using a random sampling method, with a target sample size of 200 firms and 600 employees. Ultimately, the study achieved valid responses from 145 firms and 426 employees, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes conducted by the Korea Labor Institute (2024).
In the same survey, when asked whether AI helps reduce their physical labour intensity, 71% of employees disagreed, while only 29% agreed (Figure 2.16). At the same time, when asked whether AI helps reduce their mental stress, 56% of employees disagreed, while 44% agreed (Figure 2.16).
In conclusion, the findings suggest that AI utilisation does not clearly lead to a reduction in physical labour intensity, and its impact on mental stress is mixed. While AI may substitute for certain cognitive tasks, many workers – particularly those in smaller firms and the manufacturing sector – reported no noticeable decrease in either physical or mental burden. One possible explanation is that AI adoption in Korea is still in its early stages, and its potential to ease work intensity has not yet fully materialised. Moreover, although AI can assist with cognitive processes, the need for human intervention in producing final outputs may diminish its stress-reducing effects. Overall, these results imply that while AI may enhance efficiency, its contribution to improving workers’ well-being is still uneven and inconclusive. Evidence from a case study of a game character designed in Korea confirms that the impact of AI on job quality is not black or white (Box 2.8).
Percentage of employees agreeing that AI helps reduce: (i) physical labour intensity; and (ii) mental stress
Note: The survey targeted individual employees who use AI in firms operating within four industries classified under the Korean Standard Industrial Classification (KSIC): Manufacturing, Information and Communication, Professional, Scientific and Technical Services, and Healthcare. Employees using AI provided the survey responses. The survey was conducted over a two‑month period, from 20 October 2024 to 31 December 2024. The sample was drawn using a random sampling method, with a target sample size of 600. Ultimately, the study achieved valid responses from 426 employees, whose data were incorporated into the final analysis.
Source: Survey on AI Utilisation and Labour Market Changes Conducted by the Korea Labor Institute (2024).
A character designer at a gaming company in Korea began using AI tools such as Mid-Journey and Stable Diffusion to assist with character creation. Prior to adopting AI, it took more than ten days to develop a character from an initial concept, primarily due to the time‑intensive manual sketching process. With AI, the production time has been reduced to one or two days, as the need for repetitive sketching has diminished. However, new forms of labour have emerged. The designer must now engage in prompt engineering and refine outputs through repeated trial and error to achieve the desired result. Additional stress arises from the pressure to ensure that AI-generated characters are indistinguishable from manually created ones, given the market’s growing awareness of AI use. As a result, despite the decrease in physical workload, the designer reports no clear reduction in mental stress due to the unfamiliar nature of these new tasks.
AI may be driving increases in non-standard employment
Non-standard forms of employment refer to different employment arrangements that deviate from standard employment. They include temporary employment; part-time and on-call work; temporary agency work and other multiparty employment relationships; as well as disguised employment and dependent self-employment. Non-standard employment features prominently on digital labour platforms (ILO, n.d.[62]). AI technologies are contributing to the fast growth of the platform economy. Platforms use AI-driven algorithms to match workers with tasks that align with their skills and preferences, automate administrative functions such as invoicing, implement dynamic pricing models, and optimise logistics.
Many workers – especially youth – are drawn to non-standard jobs due to low entry barriers, flexible work schedules and autonomy. In Korea, employment insurance data, which captures employees covered by employment insurance, show a significant decline in youth (i.e. aged 20‑39) employment between 2018 and 2023. While a 6% decline in the youth population accounts for part of the decline in youth employment, the remaining unexplained decline may indicate a shift of young workers from regular to non-regular employment (e.g. gig work such as delivery services).9
The changes in the employment status of 20‑39 year‑olds can be attributed to both voluntary and involuntary factors. Voluntary factors may include shifts in job preferences, for instance, platform jobs can offer substantial flexibility in working hours and days, along with competitive hourly wages (Lim, 2022[63]). Involuntary factors are more likely to reflect structural changes in the labour market. The advent of AI and robots may reduce the availability of regular jobs for which young people are qualified (Dauth et al., 2021[64]). Labour market rigidity might also make it difficult to dismiss employees before the retirement age. This discourages firms to hire, particularly younger workers who are further from retirement and thus represent long-term commitments (see Box 2.9).
Non-regular employment can meet the needs of some workers and raise labour force participation, but many non-regular jobs often lack adequate social protection. While regular employment in Korea offers strong job security and comprehensive benefits, non-regular workers face higher separation and layoff rates, lower wages – on average just over half those of regular workers – and limited social benefits coverage in pension, health and employment insurance (Tam and Xu, 2024[65]). To enhance flexibility, firms have increasingly relied on non-regular workers (Schauer, 2018[66]; Jones, 2005[67]), in particular Korea has the highest share of temporary employment among OECD countries (27% in 2024).10
Relaxing employment protection for regular workers and increasing the coverage of the social safety net for non-regular workers would help limit the extent of labour market dualism (Jones, 2005[67]; Tam and Xu, 2024[65]). Such reforms would foster a more inclusive labour market, support workforce adaptation to AI-driven transformations, and enhance both employment outcomes and firm-level productivity. Without these changes, the continued erosion in demand for regular employment could further undermine job quality.
AI is increasingly capable of automating tasks in high-skilled occupations, yet low-skilled jobs remain the most at risk of automation (Lassébie and Quintini, 2022[5]) and of job loss. Factors such as education, skills, gender, and firm size influence access to AI opportunities and its benefits, raising concerns about widening inequalities. At the same time, AI can help bridge productivity gaps between low- and high-skilled workers within occupations by capturing and disseminating best practices. While AI could help address human biases by incorporating bias detection and mitigation strategies, not carefully designed systems risk reinforcing discrimination. AI-powered solutions can also support people with disability in the workforce. Ensuring that AI fosters inclusiveness requires addressing its risks while expanding access to its opportunities.
Across OECD countries, access to AI opportunities is uneven
The introduction of AI in the workplace holds potential to improve both employment prospects and job quality. However, the benefits of AI adoption may not be distributed evenly across different groups of workers and firms. Factors such as education, skills, gender, and firm size can affect who benefits from AI advancements and who may be left behind. The uneven access to AI opportunities could prevent the benefits of AI from being broadly and fairly shared.
Workers not only vary in the extent to which they are exposed to AI, but also in their ability to adapt to and benefit from new technologies, thus the impact of AI need not be uniform across different socio-demographic groups.
High-income and high-skilled workers benefit the most from AI, while low-skilled workers may lose out
The impact of AI on employment growth has been found to be significant and positive for high-income and high-skilled occupations, and for jobs where computer use is high (Felten, Raj and Seamans, 2019[18]; Albanesi et al., 2023[20]; Georgieff and Hyee, 2021[6]). Conversely, the impact for low-skilled workers has in some cases been found to be negative. There is some evidence of an up-skilling effect, with more exposed firms reducing their employment of blue‑collar workers and hiring more high-skilled, white‑collar workers (Engberg et al., 2024[68]). Bonfiglioli et al. (2025[17]) find that AI adoption reduces employment among low-skilled and production workers, while increasing employment for workers at the top of the wage distribution and for those in STEM occupations. Similarly, Huang (2024[16]) finds that AI exposure negatively affects employment in the manufacturing and low-skill services sectors, middle‑skill workers, non-STEM occupations, and individuals at the two ends of the age distribution.
As far as wages are concerned, high-income and high-skilled occupations, as well as jobs with high computer use, tend to experience positive effects on wage growth associated with AI exposure, while lower income and lower skilled workers do not seem to benefit in the same way, or less so (Felten, Raj and Seamans, 2019[18]; Fossen and Sorgner, 2022[55]). These findings suggest that AI adoption risks increasing inequalities between groups.
These differences between workers may derive from the crucial role education plays in enhancing the ability to learn new information, making highly educated workers better equipped to adapt to new technologies (Fossen and Sorgner, 2022[55]). More educated workers are also more likely to possess the skills required to interact with, develop, and maintain AI systems. Looking at the AI workforce, for example, over 60% has at least a tertiary degree, and this figure rises to 80% within the 10 top occupations most demanding AI skills (Green and Lamby, 2023[21]). AI users are also more likely to be younger and more educated than non-users (Lane, Williams and Broecke, 2023[32]). These trends are not surprising given that 41% of low-educated adults and 33.7% of older adults lack basic proficiency in using information and communications technology (OECD, 2019[36]), making them less prepared to engage with AI systems in the workplace. Without targeted training, these workers are likely to be excluded from the benefits of AI adoption.11
Gender disparities limit women’s access to AI opportunities
The AI workforce and AI users are also predominantly male. Women constitute only 35% of the AI workforce, compared to 53% of those with tertiary education (Green and Lamby, 2023[21]). Additionally, in the OECD AI surveys of employers and workers, 41% of male workers reported using AI, compared to just 29% of female workers (Lane, Williams and Broecke, 2023[32]). Similarly, in a recent Danish study, female workers were 20 p.p. less likely to say they had used ChatGPT than male workers in the same occupation (Humlum and Vestergaard, 2024[69]). This gender disparity limits women’s access to AI-related employment opportunities and to productivity-enhancing AI tools in the workplace, and could reflect broader digital skill divides, as well as women’s lower participation in science‑related tertiary education. Although women and men across OECD countries tend to display similar levels of basic digital skills (OECD, 2019[36]), men on average tend to have more advanced digital skills like coding (European Commission, 2018[70]). Among 16‑24 year‑olds in OECD countries, 20.1% of men can write code in a programming language, while the rate for women is only 9.9% (OECD, 2023[71]). Similarly, the share of female graduates in the field of Information and Communication Technologies (ICTs) is 22.6%, while it is of 32.5% in the field of Science, Technology, Engineering and Mathematics (STEM) (OECD, 2021[72]).
Gender differences in interests, aptitudes and aspirations widen with age, suggesting that, at some point, girls are discouraged from scientific careers (OECD, 2024[73]). At age 15, boys and girls perform similarly in science and mathematics tests, however, only 1% of girls report that they want to work in ICT-related occupations, compared to 8% of boys (OECD, 2019[74]). The lack of female role models in STEM has been shown to have an impact on girls’ subject choices in school and university (Bottia et al., 2015[75]; Breda et al., 2020[76]). Additionally, discrimination in recruitment processes within technology and STEM fields results in women being less likely to gain employment, even when they possess the necessary skills (Friedmann and Efrat-Treister, 2023[77]).
Large firms are ahead of SMEs in AI adoption
The polarised adoption of AI, primarily by larger firms, combined with AI’s ability to reinforce existing inequalities, risks widening the gap between larger and smaller firms, and consequently between their workers. Data from Information and Communication Technology (ICT) surveys carried out by National Statistical Offices, reveal that AI adoption increases with firm size (see Box 2.1). Regression analysis further supports this, showing that larger firms are more likely to adopt AI, even after controlling for factors such as firm age and sector of activity (Calvino and Fontanelli, 2023[78]). Additionally, the evidence highlights the key role of complementary assets in AI adoption. AI use is more likely in presence of digital infrastructure and other digital capabilities, such as cloud computing services, other digital technologies, ultra-fast broadband connection, and higher ICT skills. The selection of larger firms into AI can be attributed to their greater resources, which enable them to face the high fixed costs associated with AI adoption and to acquire the complementary assets, related to firm digitalisation and human capital, needed to fully leverage the potential of AI (Calvino and Fontanelli, 2023[78]).
However, cost is not the only barrier to AI adoption, as evidenced by the fact that even the adoption of generative AI (which tends to be lower cost) increases with firm size (Figure 2.17). In Korea, however, medium-sized firms are less likely to use generative AI than small-sized firms, but not less likely than micro- firms (OECD, 2025[12]).
Note: The total reflects the combined results from all countries participating in the survey (Austria, Canada, Germany, Ireland, Japan, Korea, the United Kingdom).
Source: OECD (2025[12]), Microdata from the OECD SME Survey on Generative AI.
The three most commonly mentioned reasons for non-adoption of generative AI by SMEs were that: (i) generative AI is not suited to the type of work the company does (57%), (ii) concerns about copyright, legal, or regulatory issues (54%); and (iii) concerns about how the information fed into the system is handled (53%). The most reported barrier to use of generative AI in Korean SMEs was a lack of skills among employees (Figure 2.18) (OECD, 2025[12]).
Note: The total reflects the combined results from all countries participating in the survey (Austria, Canada, Germany, Ireland, Japan, Korea, the United Kingdom).
Source: OECD (2025[12]), Microdata from the OECD SME Survey on Generative AI.
Equalisation of performance within occupations
Despite little indication that AI has affected wage inequality between occupations so far, AI might be associated with lower wage inequality within occupations. AI systems demonstrate potential to enhance productivity in certain tasks, particularly for less experienced and lower-performing workers, by capturing and disseminating the effective practices of top performers, reducing performance inequality. However, for complex tasks, these tools can sometimes be counterproductive, especially for workers who struggle to identify errors in AI-generated outputs or who blindly adopt its output without careful review.
When a job’s core skill involves making predictions based on data patterns, and AI enhances the accuracy of these predictions, AI could help close the productivity gap between high- and low-skilled workers. In the context of taxi drivers, for example, an AI system assisting drivers with finding customers, by suggesting routes along which the demand is predicted to be high, narrowed the productivity gap between high- and low-productivity drivers by 14% (Kanazawa et al., 2022[79]).
Several other experiments and empirical studies focussing specifically on generative AI suggest that these tools can improve the performance of the least experienced or lowest performing workers the most. Peng et al. (2023[80]) found that GitHub Copilot, an AI tool that assists developers by providing code suggestions, helped programmers finish their task 56% faster, with the greatest gains among less experienced and older programmers. Noy and Zhang (2023[81]) showed that ChatGPT increased the productivity of college‑educated professionals in writing tasks, reducing average task time by 40% and increasing output quality by 18%, with the largest benefits for lower-performing participants. Brynjolfsson, Li and Raymond (2023[82]) analysed a generative AI-based conversational assistant, which provided real-time response suggestions to agents, while they maintained discretion to accept or ignore the suggestions. Access to the tool increased productivity (measured by issues resolved per hour) by 14% on average, including a 34% improvement for novice and low-skilled workers but with minimal impact on experienced and highly skilled workers. Dell’Acqua et al. (2023[83]) conducted a field experiment on AI’s performance in complex consulting tasks, finding that for tasks within the frontier of AI capabilities, consultants using AI were significantly more productive, both in terms of quantity and quality of the output. All consultants benefited, but those below the average performance threshold benefited more.
Analysing data across 19 OECD countries, Georgieff (2024[84]) finds some evidence that AI may be associated with lower wage inequality within occupations. While these findings are very preliminary and need to be interpreted with caution, they are consistent with the idea that AI can reduce productivity differentials between workers.
Nonetheless, researchers warn that Large Language Models can generate plausible yet incorrect results that are difficult to detect, particularly for less experienced workers. The strong performance of such tools in tasks like text summarisation, combined with confident delivery, may create a false sense of reliability across other areas. Yet, even seemingly simple tasks, such as basic mathematics, can challenge some AI models. Identifying AI’s strengths and limitations requires experience, but even experts can struggle with this distinction. Dell’Acqua et al. (2023[83]) found that, for tasks outside the frontier of AI capabilities, humans relied too much on the AI and were more likely to make mistakes when using it. Professionals who had a negative performance when using AI tended to blindly adopt its output and interrogate it less. Kabir et al. (2024[85]) analysed ChatGPT’s response to 517 programming questions, finding that 52% contained errors – yet users overlooked mistakes 39% of the time due to AI’s comprehensive and well-articulated answers. Kreitmeir and Raschky (2024[86]) exploited Italy’s sudden ChatGPT ban as a natural experiment, showing that the short-term lack of access resulted in an increase in output quantity and quality for less experienced users, but reduced productivity in routine tasks for experienced users. The authors suggest that, for complex tasks such as coding, less experienced workers may struggle to detect and correct errors in the AI-generated content but might continue relying on it due to the high cost of acquiring skills independently.
Depending on their design, algorithms can either codify or reduce bias
To effectively use AI while promoting diversity and inclusiveness, it is essential to address the potential biases that can emerge both at the data or input level and at the system level. Algorithms, if not carefully designed and implemented, can perpetuate existing human biases, leading to discrimination on a larger scale. This occurs particularly when the training data reflects historical biases against underrepresented groups due to existing discrimination in the labour market. However, AI models could also help address human bias by incorporating bias detection and mitigation strategies.
Bias at the system level is introduced through various decisions made during the development of AI systems, including the selection of variables, the decision on how to measure them, and the choice of data on which the system is trained. These decisions are often influenced by the developers’ own experiences and priors, which highlights the importance of ensuring diversity within the AI tech industry. A diverse and representative AI workforce is more likely to recognise and address these biases, leading to the development of fairer and more inclusive AI systems (Salvi del Pero, Wyckoff and Vourc’h, 2022[87]; Green and Lamby, 2023[21]).
Moreover, bias at the input or data level often derives from the use of historical data that is already biased, non-representative samples, or data that is incomplete, incorrect, or outdated. Research, such as correspondence experiments in economics, has shown that human decision making is also biased. These studies, which involve sending to real job openings fictitious job applications differing only in a randomly assigned characteristic, reveal discrimination against ethnic minorities, pregnant women, mothers, transgender people, Muslims, and people with disabilities (Bertrand and Duflo, 2017[88]; Baert, 2017[89]). Since applicants from underrepresented groups are less likely to be hired for high-income jobs due to existing biases, they are underrepresented in training datasets. As a result, AI models trained on such data may disproportionately favour historically hired groups, reinforcing discriminatory practices and reducing diversity compared to human-led hiring processes (Li, Raymond and Bergman, 2020[90]). While algorithms are not responsible for societal biases, they can replicate, ingrain, and/or amplify them at scale. This automated discrimination is often more abstract and harder to detect than traditional forms, complicating enforcement of non-discrimination laws (Salvi del Pero, Wyckoff and Vourc’h, 2022[87]).
Nonetheless, if designed and implemented well, AI models could be used to address and mitigate human bias. For instance, in 2018, LinkedIn introduced an AI feature that ensured that top search results seen by recruiters had a gender breakdown more representative of the potential applicant pool (Bogen and Rieke, 2018[91]; Chan, 2018[92]). Some firms are dedicated to developing diversity and inclusion technologies, which aim to reduce bias in AI systems through innovative techniques. One approach, known as “post-processing,” involves setting quotas to ensure that the recommendations made by the algorithm include a balanced representation from various population sub-groups. Another method involves training the algorithm with artificial profiles, thereby increasing the likelihood that candidates from underrepresented groups will be selected (Broecke, 2023[93]).
AI-powered tools for workers with disabilities
AI has the potential to support people with disability in the labour market. This technology has been implemented in numerous tools that can either reduce barriers by providing workarounds for people with disability (such as live captioning for deaf individuals), or that address disabilities directly (such as AI-powered prosthetics). Other tools make content and workplaces more accessible. However, significant challenges remain in funding, accessibility, and user engagement.
Building on desk research and interviews with over 70 stakeholders, an OECD study (Touzet, 2023[94]) identified 142 AI-powered solutions that can support people with disabilities in the workforce, 75% of which would not exist in their present form without AI. More than half of the identified tools are disability-centred, designed to either directly address disabilities or provide workarounds. Examples include AI-powered hearing aids, live captioning algorithms for deaf or hard-of-hearing individuals, image recognition solutions allowing blind or low-vision individuals to hear descriptions of their surroundings, self-driving wheelchairs, and speech-to-text solutions for people with dysarthric speech or who cannot type on a keyboard. Other tools identified in the study focus on making environments more inclusive by adapting content and workplaces rather than requiring persons with disabilities to adjust. Examples include AI powered indoor positioning systems to make buildings accessible to blind and low-vision individuals, as well as natural language processing applications that translate written content into plain language, improving accessibility for neurodiverse individuals.
Furthermore, AI is enhancing prosthetics by interpreting the user’s intentions based on physiological data and brain signals, enabling robotic parts to perform tasks like opening and closing the hand, moving individual fingers, and navigating stairs and ramps. AI can also quantify external touch and sensory input, allowing prostheses to convey these sensations to the brain. Additionally, AI-powered voice assistants can help people carry out routine daily actions through voice commands, such as turning lights on and off, regulating room temperature, and unlocking doors (Pancholi, Wachs and Duerstock, 2024[95]). AI’s versatility allows for broader applications compared to traditional assistive technologies, offering greater customisation and embedding accessibility features into mainstream products.
However, the development and adoption of these technologies face several challenges. One major barrier is the scarcity of funding for research and development (R&D) as both public and private investment in disability-focussed innovations remain scarce, often leaving promising tools at the prototype stage. Commercialisation also presents hurdles: direct sales to users can create inequities, while integrating public reimbursement pathways is often too complex for small firms. Employer adoption is further constrained by limited awareness of accessibility needs. This limits the reach of AI-powered solutions, restricting access to those who can afford them and are informed about their existence. Additionally, the lack of user engagement in the development process can lead to irrelevant or impractical solutions. Lastly, low levels of IT literacy among some end users and issues with interoperability between new AI solutions and existing assistive technologies further hinder adoption (Touzet, 2023[94]).
In terms of the impact of AI on inclusiveness, a significant gap between large firms and SMEs in Korea in terms of both productivity and AI adoption has already been documented in this report. In addition, the negative impact of traditional AI on regular, full-time employment growth appears to be concentrated among younger workers, low- to medium-skilled workers and women – although for the latter, as well as for high-skilled workers, higher exposure to generative AI is associated with higher employment growth. Generative AI is associated with higher wage growth for men and high-skilled workers, while traditional AI is associated with higher wage growth for older workers and high-skilled workers. By contrast, traditional AI appears to reduce wage growth for low-skilled workers. Once again, however, it needs to be remembered that these findings apply to full-time, permanent employees only, and the findings for non-standard workers may be different. Further details for these findings are provided in Box 2.9.
This box breaks the analysis of Box 2.4 down by gender, age and skill level.
Gender-based heterogeneity was found in both traditional AI and generative AI. Table 2.3 shows that for women, traditional AI (Webb index) is associated with a statistically significant negative effect on employment growth, while generative AI (Felten index) exhibits a statistically significant positive effect. In contrast to its impact on employment, there is a positive effect of generative AI on wage growth for male employees (Table 2.4). This finding is particularly noteworthy given that, on average, women are more exposed to generative AI. This apparent discrepancy stems from the fact that men are disproportionately represented in occupations situated within the higher percentiles of generative AI exposure.
In the age subgroup analysis, “Young” refers to individuals aged 39 and below,1 “Middle” to those aged 40‑49, and “Old” to those aged 50 and above. The negative association of AI with employment growth was statistically significant for traditional AI only among the younger cohort, with no observed effects in other age groups (Table 2.3). This finding suggests that the reduced labour demand stemming from AI exposure might have operated more by curtailing new hiring than through layoffs of existing employees. In Korea’s relatively rigid labour market, where layoffs are challenging, firms may be hesitant to increase new hires due to difficulties in flexible workforce adjustment following the introduction of AI technology. This phenomenon is similar to that observed in German manufacturing following the introduction of industrial robots (Dauth et al., 2021[64]). A statistically significant and positive relationship between AI and wage growth is only observed for older workers and traditional AI (Table 2.4), which might be because older workers tend to be more exposed to traditional AI, and less so to generative AI. This points to a more nuanced dynamic – while traditional AI may dampen employment growth among younger cohorts, it appears to support wage growth for older workers already in employment, particularly in roles characterised by higher levels of AI exposure.
Finally, traditional AI appears to be associated with lower employment growth among low-skilled and medium-skilled workers, while generative AI is associated with higher employment growth among high-skilled workers (Table 2.3). Both kinds of AI are also associated with higher wage growth for high-skilled workers only, while traditional AI is associated with lower wage growth for low-skilled workers (Table 2.4). This result aligns with the findings of Bonfiglioli et al. (2025[17]), who posited a positive impact of AI adoption on STEM occupations and high-wage workers. The divergence between low- and high-skilled workers is likely driven by distinct market dynamics. In the case of generative AI, a rapid increase in demand for highly skilled labour, combining with relatively inelastic supply, has resulted in a statistically robust positive effect on wage growth. By contrast, traditional AI appears to exert a displacement effect on lower-skilled labour, contributing to downward pressure on wage increases for these groups.
Once again, however, it needs to be remembered that these findings apply to full-time, permanent employees only, and the findings for non-standard workers may be different.
Note: Standard errors are shown in parentheses, and statistical significance is indicated as follows: *** 1%, ** 5%, * 10%.
1. In general, the OECD defines “youth” as those individuals aged 25 and under. However, in the context of Korea, the labour market participation rate among those under 20 is extremely low, and the target group for youth policies typically ranges from 20 to 39 years of age. Therefore, the age bracket for youth has been defined as 20 to 39 years in this report.
Source: Employment insurance data and information provided by Webb (2020[43]), “The Impact of Artificial Intelligence on the Labor Market”, https://www.michaelwebb.co/webb_ai.pdf; and Felten et al. (2023[10]), “How will Language Modelers like ChatGPT Affect Occupations and Industries?”, https://doi.org/10.48550/arXiv.2303.01157 compiled by the authors.
References
[25] Acemoglu, D. (2024), The Simple Macroeconomics of AI, National Bureau of Economic Research, Cambridge, MA, https://doi.org/10.3386/w32487.
[15] Acemoglu, D. et al. (2022), “Artificial Intelligence and Jobs: Evidence from Online Vacancies”, Journal of Labor Economics, Vol. 40/S1, https://doi.org/10.1086/718327.
[13] Acemoglu, D. and P. Restrepo (2019), “Automation and New Tasks: How Technology Displaces and Reinstates Labor”, Journal of Economic Perspectives, Vol. 33/2, pp. 3-30, https://doi.org/10.1257/JEP.33.2.3.
[20] Albanesi, S. et al. (2023), “New Technologies and Jobs in Europe”, http://www.nber.org/papers/w31357 (accessed on 21 June 2024).
[27] Alekseeva, L. et al. (2021), “The demand for AI skills in the labor market”, Labour Economics, Vol. 71, p. 102002, https://doi.org/10.1016/J.LABECO.2021.102002.
[23] Anderson, B. and E. Sutherland (2024), “Collective action for responsible AI in health”, OECD Artificial Intelligence Papers, No. 10, OECD Publishing, Paris, https://doi.org/10.1787/f2050177-en.
[1] Autor, D., F. Levy and R. Murnane (2003), “The Skill Content of Recent Technological Change: An Empirical Exploration”, The Quarterly Journal of Economics, https://doi.org/10.1162/003355303322552801.
[89] Baert, S. (2017), “Hiring Discrimination: An Overview of (Almost) All Correspondence Experiments Since 2005”, http://www.iza.org (accessed on 29 August 2024).
[88] Bertrand, M. and E. Duflo (2017), “Field Experiments on Discrimination”, in Handbook of Economic Field Experiments, North-Holland, https://doi.org/10.1016/BS.HEFE.2016.08.004.
[31] Bessen, J. et al. (2023), “The Business of AI Startups”, SSRN Electronic Journal, https://doi.org/10.2139/SSRN.3293275.
[91] Bogen, M. and A. Rieke (2018), Help Wanted: An Examination of Hiring Algorithms, Equity, and Bias, Upturn.
[17] Bonfiglioli, A. et al. (2025), “Artificial intelligence and jobs: evidence from US commuting zones”, Economic Policy, Vol. 40/121, pp. 145-194, https://doi.org/10.1093/EPOLIC/EIAE059.
[75] Bottia, M. et al. (2015), “Growing the roots of STEM majors: Female math and science high school faculty and the participation of students in STEM”, Economics of Education Review, Vol. 45, pp. 14-27, https://doi.org/10.1016/j.econedurev.2015.01.002.
[76] Breda, T. et al. (2020), “Do Female Role Models Reduce the Gender Gap in Science? Evidence from French High Schools”, https://www.iza.org/publications/dp/13163/do-female-role-models-reduce-the-gender-gap-in-science-evidence-from-french-high-schools (accessed on 28 August 2024).
[59] Brione, P. (2020), My boss the algorithm: Research An ethical look at algorithms in the workplace, ACAS.
[93] Broecke, S. (2023), “Artificial intelligence and labour market matching”, OECD Social, Employment and Migration Working Papers, No. 284, OECD Publishing, Paris, https://doi.org/10.1787/2b440821-en.
[82] Brynjolfsson, E., D. Li and L. Raymond (2023), “Generative AI at Work”, National Bureau of Economic Research Working Paper Series, Vol. No. 31161, https://doi.org/10.3386/W31161.
[19] Brynjolfsson, E., D. Rock and C. Syverson (2017), “Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics”, National Bureau of Economic Research Working Paper Series, https://doi.org/10.3386/w24001 (accessed on 25 July 2024).
[78] Calvino, F. and L. Fontanelli (2023), “A portrait of AI adopters across countries: Firm characteristics, assets’ complementarities and productivity”, OECD Science, Technology and Industry Working Papers, No. 2023/02, OECD Publishing, Paris, https://doi.org/10.1787/0fb79bb9-en.
[53] Cazes, S., A. Hijzen and A. Saint-Martin (2015), “Measuring and Assessing Job Quality: The OECD Job Quality Framework”, OECD Social, Employment and Migration Working Papers, No. 174, OECD Publishing, Paris, https://doi.org/10.1787/5jrp02kjw1mr-en.
[42] Chang, J. et al. (2024), Changing work processes and employment in the age of AI, Korea Labor Institute, https://www.kli.re.kr/board.es?mid=a20403000000&bid=0071&act=view&list_no=145216&tag=&nPage=2 (accessed on 13 June 2025).
[92] Chan, R. (2018), LinkedIn is using AI to make recruiting diverse candidates a no-brainer, Business Insider, https://www.businessinsider.com/linkedin-new-ai-feature-increase-diversity-hiring-2018-10 (accessed on 29 August 2024).
[64] Dauth, W. et al. (2021), “The Adjustment of Labor Markets to Robots”, Journal of the European Economic Association, Vol. 19/6, pp. 3104-3153, https://doi.org/10.1093/JEEA/JVAB012.
[83] Dell’Acqua, F. et al. (2023), “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality”, SSRN Electronic Journal, https://doi.org/10.2139/SSRN.4573321.
[9] Eloundou, T. et al. (2023), “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models”, https://arxiv.org/abs/2303.10130v5 (accessed on 4 September 2024).
[68] Engberg, E. et al. (2024), “AI Unboxed and Jobs: A Novel Measure and Firm-Level Evidence from Three Countries”, SSRN Electronic Journal, https://doi.org/10.2139/SSRN.4691172.
[70] European Commission (2018), Women in the digital age - Executive summary, Publications Office, https://doi.org/10.2759/517222.
[10] Felten, E., M. Raj and R. Seamans (2023), “How will Language Modelers like ChatGPT Affect Occupations and Industries?”, SSRN Electronic Journal, https://doi.org/10.2139/SSRN.4375268.
[7] Felten, E., M. Raj and R. Seamans (2021), “Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses”, Strategic Management Journal, Vol. 42/12, pp. 2195-2217, https://doi.org/10.1002/SMJ.3286.
[18] Felten, E., M. Raj and R. Seamans (2019), “The Occupational Impact of Artificial Intelligence: Labor, Skills, and Polarization”, SSRN Electronic Journal, https://doi.org/10.2139/SSRN.3368605.
[47] Felten, E., M. Raj and R. Seamans (2018), “A Method to Link Advances in Artificial Intelligence to Occupational Abilities”, AEA Papers and Proceedings, Vol. 108, pp. 54-57, https://doi.org/10.1257/PANDP.20181021.
[54] Filippucci, F., P. Gal and M. Schief (2024), “Miracle or Myth? Assessing the macroeconomic productivity gains from Artificial Intelligence”, OECD Artificial Intelligence Papers, No. 29, OECD Publishing, Paris, https://doi.org/10.1787/b524a072-en.
[55] Fossen, F. and A. Sorgner (2022), “New digital technologies and heterogeneous wage and employment dynamics in the United States: Evidence from individual-level data”, Technological Forecasting and Social Change, Vol. 175, p. 121381, https://doi.org/10.1016/J.TECHFORE.2021.121381.
[77] Friedmann, E. and D. Efrat-Treister (2023), “Gender Bias in Stem Hiring: Implicit In-Group Gender Favoritism Among Men Managers”, Gender and Society, Vol. 37/1, pp. 32-64, https://doi.org/10.1177/08912432221137910.
[84] Georgieff, A. (2024), “Artificial intelligence and wage inequality”, OECD Artificial Intelligence Papers, No. 13, OECD Publishing, Paris, https://doi.org/10.1787/bf98a45c-en.
[6] Georgieff, A. and R. Hyee (2021), “Artificial intelligence and employment : New cross-country evidence”, OECD Social, Employment and Migration Working Papers, No. 265, OECD Publishing, Paris, https://doi.org/10.1787/c2c1d276-en.
[34] Green, A. (2024), “Artificial intelligence and the changing demand for skills in the labour market”, OECD Artificial Intelligence Papers, No. 14, OECD Publishing, Paris, https://doi.org/10.1787/88684e36-en.
[21] Green, A. and L. Lamby (2023), “The supply, demand and characteristics of the AI workforce across OECD countries”, OECD Social, Employment and Migration Working Papers, Vol. No. 287, https://doi.org/10.1787/bb17314a-en.
[3] Guo, B. et al. (2023), “How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection”, https://arxiv.org/abs/2301.07597v1 (accessed on 25 July 2024).
[39] Han, J. (2023), 인공지능으로 인한 노동시장의 변화와 정책방향 (The Impact of Artificial Intelligence on the Labor Market and Policy Implications), Korea Development Institute (KDI), https://doi.org/10.2139/SSRN.4890349.
[16] Huang, Y. (2024), “The Labor Market Impact of Artificial Intelligence: Evidence from US Regions”, IMF Working Papers, Vol. 2024/199, p. 1, https://doi.org/10.5089/9798400288555.001.
[69] Humlum, A. and E. Vestergaard (2024), “The Adoption of ChatGPT”, SSRN Electronic Journal, https://doi.org/10.2139/SSRN.4827166.
[62] ILO (n.d.), Non-standard forms of employment | International Labour Organization, https://www.ilo.org/topics-and-sectors/non-standard-forms-employment (accessed on 12 June 2025).
[67] Jones, R. (2005), “The Labour Market in Korea: Enhancing Flexibility and Raising Participation”, OECD Economics Department Working Papers, No. 469, OECD Publishing, Paris, https://doi.org/10.1787/548701741108.
[50] Jung, J. (2018), “Domestic and overseas doctorates and their academic entry-level jobs in South Korea”, Asian Education and Development Studies, Vol. 7/2, pp. 205-222, https://doi.org/10.1108/AEDS-07-2017-0070.
[85] Kabir, S. et al. (2024), “Is Stack Overflow Obsolete? An Empirical Study of the Characteristics of ChatGPT Answers to Stack Overflow Questions”, Conference on Human Factors in Computing Systems - Proceedings, https://doi.org/10.1145/3613904.3642596.
[79] Kanazawa, K. et al. (2022), “AI, Skill, and Productivity: The Case of Taxi Drivers”, https://docs.iza.org/dp15677.pdf (accessed on 22 August 2024).
[46] Kim, S., S. Koh and Y. Cho (2014), Dynamic Effects of Technological Advancement on the Labor Market, Korea Labor Institute, https://www.kli.re.kr/kli_eng/rschRptpAuthorList.es?sch_type=aut_nm&sch_keyword=%20Seon%20Koh&mid=a20101000000&nPage=1 (accessed on 13 June 2025).
[44] Korea Employment Information Service (2020), 2020 Korea Dictionary of Occupations, The Ministry of Employment and Labor and Korea Employment Information Service.
[40] KOSIS (2025), Survey of Business Activities 2023 - Technology Development and Utilization: Artificial Intelligence, https://kosis.kr/statHtml/statHtml.do?sso=ok&returnurl=https%3A%2F%2Fkosis.kr%3A443%2FstatHtml%2FstatHtml.do%3Flist_id%3DJ2_7_300_009%26obj_var_id%3D%26seqNo%3D%26tblId%3DDT_1EP1337_1%26vw_cd%3DMT_ZTITLE%26orgId%3D101%26path%3D%252FstatisticsList%252FstatisticsListIndex.do%26conn_path%3DMT_ZTITLE%26itm_id%3D%26lang_mode%3Dko%26scrId%3D%26 (accessed on 18 June 2025).
[86] Kreitmeir, D. and P. Raschky (2024), “The Heterogeneous Productivity Effects of Generative AI”, SSRN Electronic Journal, https://doi.org/10.2139/ssrn.4745624.
[45] Kwon, J. (2022), Characteristics and Mobility Factors of Low-Wage Workers Based on Employment Insurance Administrative Data, Korea Employment Information Service.
[8] Lane, M. (2024), “Who will be the workers most affected by AI?: A closer look at the impact of AI on women, low-skilled workers and other groups”, OECD Artificial Intelligence Papers, No. 26, OECD Publishing, Paris, https://doi.org/10.1787/14dc6f89-en.
[26] Lane, M. and A. Saint-Martin (2021), “The impact of Artificial Intelligence on the labour market: What do we know so far?”, OECD Social, Employment and Migration Working Papers, No. 256, OECD Publishing, Paris, https://doi.org/10.1787/7c895724-en.
[32] Lane, M., M. Williams and S. Broecke (2023), “The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers”, OECD Social, Employment and Migration Working Papers, No. 288, OECD Publishing, Paris, https://doi.org/10.1787/ea0a0fe1-en.
[5] Lassébie, J. and G. Quintini (2022), “What skills and abilities can automation technologies replicate and what does it mean for workers?: New evidence”, OECD Social, Employment and Migration Working Papers, No. 282, OECD Publishing, Paris, https://doi.org/10.1787/646aad77-en.
[90] Li, D., L. Raymond and P. Bergman (2020), “Hiring as Exploration”, National Bureau of Economic Research Working Paper Series, Vol. No. 27736, https://doi.org/10.3386/w27736.
[63] Lim, J. (2022), “Determinants of Wage for Web-based Platform Workers: In perspective of evaluation by previous employers”, Journal of Digital Convergence, Vol. 20/4, pp. 1-14, https://doi.org/10.14400/JDC.2022.20.4.001.
[24] Machin, J. (2024), AI can help people keep physical jobs for longer, https://www.bbc.com/news/uk-england-68969349 (accessed on 4 September 2024).
[28] Manca, F. (2023), “Six questions about the demand for artificial intelligence skills in labour markets”, OECD Social, Employment and Migration Working Papers, No. 286, OECD Publishing, Paris, https://doi.org/10.1787/ac1bebf0-en.
[51] Maslej, N. et al. (2025), The AI Index 2025 Annual Report, AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, https://hai.stanford.edu/ai-index/2025-ai-index-report (accessed on 30 April 2025).
[2] Maslej, N. et al. (2024), The AI Index 2024 Annual Report, AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, https://hai.stanford.edu/ai-index/2024-ai-index-report.
[49] Matsuura, N. (2024), Robot sommeliers and baristas go to work in labor-starved South Korea - Nikkei Asia, https://asia.nikkei.com/Business/Technology/Robot-sommeliers-and-baristas-go-to-work-in-labor-starved-South-Korea (accessed on 8 August 2024).
[60] Meijer, A., L. Lorenz and M. Wessels (2021), “Algorithmization of Bureaucratic Organizations: Using a Practice Lens to Study How Context Shapes Predictive Policing Systems”, Public Administration Review, Vol. 81/5, pp. 837-846, https://doi.org/10.1111/PUAR.13391.
[14] Milanez, A. (2023), “The impact of AI on the workplace: Evidence from OECD case studies of AI implementation”, OECD Social, Employment and Migration Working Papers, No. 289, OECD Publishing, Paris, https://doi.org/10.1787/2247ce58-en.
[33] Milanez, A., A. Lemmens and C. Ruggiu (2025), “Algorithmic management in the workplace: New evidence from an OECD employer survey”, OECD Artificial Intelligence Papers, No. 31, OECD Publishing, Paris, https://doi.org/10.1787/287c13c4-en.
[61] Moon, A. (2023), The Impact of Generative AI on the Labor Market and Policy Implications, Korea Information Society Development Institute, https://library.kisdi.re.kr/$/10160/contents/4333450?checkinId=2071069&articleId=1171956 (accessed on 13 June 2025).
[56] Moore, P. (2019), “OSH and the Future of Work: Benefits and Risks of Artificial Intelligence Tools in Workplaces”, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 11581 LNCS, pp. 292-315, https://doi.org/10.1007/978-3-030-22216-1_22.
[58] Moore, V. (2018), The Threat of Physical and Psychosocial Violence and Harassment in Digitalized Work, International Labour Office, Geneva.
[41] NIA (2025), 2024 Yearbook of Enterprise Informatization Statistics, National Information Society Agency and Ministry of Science and ICT, Republic of Korea, https://eng.nia.or.kr/site/nia_eng/ex/bbs/View.do;jsessionid=34AFDD692F3F88105EC7F1B30DE3B071.7cc30ba5eea206361634?cbIdx=31975&bcIdx=27694&parentSeq=27694 (accessed on 25 April 2025).
[81] Noy, S. and W. Zhang (2023), “Experimental evidence on the productivity effects of generative artificial intelligence”, Science, Vol. 381/6654, pp. 187-192, https://doi.org/10.1126/SCIENCE.ADH2586.
[97] O*NET (2025), Abilities — Trunk Strength, https://www.onetonline.org/find/descriptor/result/1.A.3.a.4 (accessed on 29 April 2025).
[11] OECD (2025), ICT Access and Usage by Businesses, Businesses using artificial intelligence (AI), https://data-viewer.oecd.org/?chartId=931683b4-248c-406e-860e-85a464fe8a73 (accessed on 31 January 2025).
[12] OECD (2025), “Microdata from the OECD SME Survey on Generative AI”.
[52] OECD (2025), Trends in Adult Learning: New Data from the 2023 Survey of Adult Skills, Getting Skills Right, OECD Publishing, Paris, https://doi.org/10.1787/ec0624a6-en.
[73] OECD (2024), Bridging Talent Shortages in Tech: Skills-first Hiring, Micro-credentials and Inclusive Outreach, Getting Skills Right, OECD Publishing, Paris, https://doi.org/10.1787/f35da44f-en.
[48] OECD (2024), OECD Economic Surveys: Korea 2024, OECD Publishing, Paris, https://doi.org/10.1787/c243e16a-en.
[22] OECD (2024), OECD Employment Outlook 2024: The Net-Zero Transition and the Labour Market, OECD Publishing, Paris, https://doi.org/10.1787/ac8b3538-en.
[35] OECD (2024), Training Supply for the Green and AI Transitions: Equipping Workers with the Right Skills, Getting Skills Right, OECD Publishing, Paris, https://doi.org/10.1787/7600d16d-en.
[4] OECD (2023), Is Education Losing the Race with Technology?: AI’s Progress in Maths and Reading, Educational Research and Innovation, OECD Publishing, Paris, https://doi.org/10.1787/73105f99-en.
[38] OECD (2023), OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market, OECD Publishing, Paris, https://doi.org/10.1787/08785bba-en.
[71] OECD (2023), Women as a share of all 16-24 year-olds who can program | Society Indicators, https://goingdigital.oecd.org/indicator/54 (accessed on 28 August 2024).
[72] OECD (2021), OECD Data Explorer - Archive • Distribution of graduates and new entrants by field, https://data-explorer.oecd.org/vis?tenant=archive&df[ds]=DisseminateArchiveDMZ&df[id]=DF_EAG_GRAD_ENTR_FIELD&df[ag]=OECD&dq=OAVG.ENTR_SHARE_GENDER-FIELDS%2BGRAD_SHARE_GENDER-FIELDS.ISCED11_5T8.._T.F05T07%2BF06&lom=LASTNPERIODS&lo=5&to[TIME_PERIOD]=false&vw=tb (accessed on 28 August 2024).
[37] OECD (2019), Getting Skills Right: Future-Ready Adult Learning Systems, Getting Skills Right, OECD Publishing, Paris, https://doi.org/10.1787/9789264311756-en.
[74] OECD (2019), PISA 2018 Results (Volume II): Where All Students Can Succeed, PISA, OECD Publishing, Paris, https://doi.org/10.1787/b5fd1b8f-en.
[36] OECD (2019), Skills Matter: Additional Results from the Survey of Adult Skills, OECD Skills Studies, OECD Publishing, Paris, https://doi.org/10.1787/1f029d8f-en.
[95] Pancholi, S., J. Wachs and B. Duerstock (2024), “Use of Artificial Intelligence Techniques to Assist Individuals with Physical Disabilities”, Annual Review of Biomedical Engineering, Vol. 26/1, pp. 1-24, https://doi.org/10.1146/ANNUREV-BIOENG-082222-012531.
[80] Peng, S. et al. (2023), “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot”.
[87] Salvi del Pero, A., P. Wyckoff and A. Vourc’h (2022), “Using Artificial Intelligence in the workplace: What are the main ethical risks?”, OECD Social, Employment and Migration Working Papers, No. 273, OECD Publishing, Paris, https://doi.org/10.1787/840a2d9f-en.
[66] Schauer, J. (2018), “Labor Market Duality in Korea”, IMF Working Papers.
[29] Squicciarini, M. and H. Nachtigall (2021), “Demand for AI skills in jobs: Evidence from online job postings”, OECD Science, Technology and Industry Working Papers, No. 2021/03, OECD Publishing, Paris, https://doi.org/10.1787/3ed32d94-en.
[65] Tam, S. and X. Xu (2024), “Advancing Labor Market Reforms in Korea”, IMF Working Papers, Vol. 2024/183, p. 1, https://doi.org/10.5089/9798400289026.001.
[57] Todolí-Signes, A. (2021), “Making algorithms safe for workers: occupational risks associated with work managed by artificial intelligence”, European Review of Labour and Research, Vol. 27/4, pp. 433-452, https://doi.org/10.1177/10242589211035040.
[94] Touzet, C. (2023), “Using AI to support people with disability in the labour market: Opportunities and challenges”, OECD Artificial Intelligence Papers, No. 7, OECD Publishing, Paris, https://doi.org/10.1787/008b32b7-en.
[43] Webb, M. (2020), “The Impact of Artificial Intelligence on the Labor Market”, Stanford University, https://www.michaelwebb.co/webb_ai.pdf (accessed on 29 April 2025).
[30] Wilson, H., P. Daugherty and N. Morini-Bianzino (2017), “The Jobs That Artificial Intelligence Will Create”, Vol. 58/4, http://mitsmr.com/2odREFJ (accessed on 21 June 2024).
[96] Zhang, D. et al. (2022), The AI Index 2022 Annual Report.
← 1. Image classification refers to the ability of machines to categorise what they see in images (Zhang et al., 2022[96]).
← 2. Visual reasoning assesses how well AI systems can reason across a combination of visual and textual data (Zhang et al., 2022[96]). AI models take an image and a textual question as input and generate a relevant answer.
← 3. Natural language inference is the task of determining whether, given a premise, a hypothesis is true (entailment), false (contradiction), or undetermined (neutral) (Zhang et al., 2022[96]).
← 4. Trunk strength is the ability to use your abdominal and lower back muscles to support part of the body repeatedly or continuously over time without “giving out” or fatiguing (O*NET[97]).
← 5. Many of these occupations remain exposed to technologies other than AI. In addition, AI is frequently embedded in such technologies (e.g. robots).
← 6. Visual commonsense reasoning (VCR) refers to the ability of AI systems to not only answer questions based on images but also reason about the logic behind their answers. Performance in VCR is measured by evaluating the machine’s ability to both select the correct answer to a question and choose the appropriate rationale behind that answer (Maslej et al., 2024[2]).
← 7. The OECD case studies of AI implementation were conducted in Austria, Canada, France, Germany, Ireland, Japan, the United Kingdom and the United States.
← 8. France, Germany, Italy, Spain.
← 9. To isolate the impact of changes in employment insurance eligibility criteria, the analysis was confined to salaried, permanent employees, with the explicit exclusion of self-employed, non-regular, and freelance workers. The sample was further refined to include only those regular employees for whom complete data were available across all primary variables, specifically gender, age, occupation, and industry.
← 10. OECD 2024 Employment by permanency of the job – Incidence http://data-explorer.oecd.org/s/24j.
← 11. With respect to youth, studies based on jobs and studies based on individuals may give opposite results. Youth come out as more at risk in studies that assess automatability at the level of tasks/jobs because they tend to hold jobs with fewer high-level cognitive tasks (career-related). However, looking at individuals directly, youth are better prepared because they have better digital skills.