The impact AI will have on the labour market, including whether the benefits will outweigh the risks and how these will be distributed, will depend on the policies, regulations and institutions countries have in place. This chapter reviews how workers can be supported through the transition, including through training, social protection, employment services and social dialogue, and also what steps can be taken to ensure no one is left behind, and promote the safe and trustworthy use in the workplace.
Artificial Intelligence and the Labour Market in Korea
Abstract
Supporting workers through the transition
As documented in Chapter 1, AI will likely result in significant change in the labour market, and workers will need help in managing the transition. Outcomes for workers are better when they are consulted about the use of AI in the workplace, and also when they have received training. For workers who will be displaced by AI, social protection combined with re‑employment services will be essential to help them find new jobs.
Most OECD countries, including Korea, already have strong institutions and systems in place for training, social protection and re‑employment services. However, AI brings new challenges which may require either a scaling-up of existing efforts, or more AI-specific solutions.
Korea has already implemented several publicly funded education and training programmes to develop AI skills, including K-Digital Training, a vocational training initiative that aims to provide high-skilled workers in digital and edge‑tech industries. To further improve the alignment of skills needs and supply, Korea should aim to: (i) strategically co‑ordinate training policies between the Ministry of Trade, Industry and Energy and the Ministry of Employment and Labour; (ii) tailor training programmes to the specific needs of SMEs; and (iii) promote more work-based learning.
AI may be able to assist in the provision and improvement of training. For instance, AI has the potential to increase training participation, including among currently underrepresented groups, by lowering some of the barriers to training that people experience, increasing motivation to train, personalising content and delivery, and offering more targeted learning pathways. Moreover, certain AI solutions for training may improve the alignment of training to labour market needs, and help in planning and delivering training policies.
In Korea, the Ministry of Employment and Labour has also begun deploying AI technologies in the provision of employment services and counselling to job seekers (Work24), thereby improving labour market matching.
Promoting an inclusive transition
AI will not affect all groups equally. Some individuals are more likely to gain, while others are at higher risk of either losing their job or experiencing changes on the job. Policymakers should aim to promote an inclusive transition, targeting training, as well as employment and social support on those who need it most. At the same time, the development and adoption of AI tools that boost the employability of underrepresented groups (e.g. those with disabilities) need to be promoted, and bias and discrimination embedded in AI tools need to be addressed.
Promoting safe and trustworthy AI
Policymakers should take steps to ensure that the AI that is used in the workplace is safe and trustworthy. This is not only to protect workers, but also to promote the adoption of such tools and reduce potential worker resistance. Specifically, and in line with the OECD AI Principles:
Worker health and safety needs to be protected. While AI may boost productivity and lower costs, this should not come at the expense of increased worker stress, anxiety or risk of accidents. Most countries have regulations that set out employers’ obligations towards employees concerning their occupational safety and health. While in theory such regulations should also cover AI, there may be gaps. Also, while most countries have product liability regulations, these might need to be adapted to the use of AI systems. Moreover, labour inspectorates may lack the knowledge and/or capacity to address new risks posed by AI.
Workers personal data and privacy should be protected. The growing integration of AI in the world of work will likely result in greater collection and analysis of data on workers and job applicants. Workers may worry about their privacy when such data is personal. They might also worry that the data are used for purposes other than for which they were intended. Increased monitoring and surveillance could also lead to stress. In many countries, data protection and privacy rules, or their implementation, may need to be strengthened. In particular, asking workers for consent to collect and use personal data may not be meaningful in the context of the workplace and the inherent power imbalance between employers and workers. In Korean workplaces, consent is deemed to be given when the union agrees with the data collection, while in workplaces without unions, the condition is met when more than half of all workers express their agreement.
There should be human agency and oversight. Automated decision making and algorithmic management practices can pose threats to workers, particularly where they impact consequential decisions such as hiring, evaluations, promotions, and terminations. Policymakers across OECD countries are taking steps to regulate automated decision making by: clarifying the circumstances under which automated decision making may or may not be used, requiring human oversight throughout the AI lifecycle as well as transparency; requiring periodic testing and validation of the AI; and giving individuals rights to: human intervention, express their opinion, to obtain an explanation, and challenge/contest decisions taken by algorithms. One of the challenges in regulating automated decision making is how to make a “human in the loop” more than just the rubberstamping of automatic decisions.
There should be transparency in the use of AI in the workplace. The ability of workers to detect risks or harms, effectively question outcomes, and exercise their rights, hinges on their awareness of their interactions with AI systems and how those systems reach their outcomes. Most AI principles that have been published in OECD countries to date underscore the importance of transparency of AI and its use, and several countries are introducing legislation requiring employers to notify individuals when they are interacting with certain forms of AI.
Workers should be entitled to meaningful information about the logic involved in AI outputs. AI systems yield outcomes that can be difficult or even impossible to explain. A lack of explainability can undermine the trust and confidence that people place in AI systems and the decisions that are informed by them. It also makes it difficult for individuals to provide informed consent to the use of such systems, or to identify and seek redress for adverse effects caused by AI systems in the workplace.
Establishing clear lines of accountability is fundamental for a trustworthy use of AI and the ability to enforce regulations. In recent years, legislators across the OECD, have made efforts to promote accountability mechanisms, such as impact assessments and/or audits of AI systems to provide evidence and assurance that they are trustworthy and safe to use.
In 2025, Korea passed a draft AI Basic Act to be implemented in 2026 which aims to promote AI development. At the same time, the AI Basic Act aims to “minimise risks and build trust”, and covers important principles such as transparency and explainability, as well as safety and reliability of AI. Korea’s approach to regulating AI is based more on principles than on detailed prescriptions and regulations, and the private sector will be encouraged to develop and implement its own mechanisms for ensuring AI safety and reliability. That being said, AI does not operate in a regulatory vacuum and existing laws, such as the Labour Standards Act and the Personal Information Protection Act, remain applicable.
Social dialogue could play an important role in making a success of AI in the workplace. OECD evidence shows that workers who have been consulted about the use of new technologies are more positive about the impact of AI on their work than those who have not. In Korea, the Economic, Social and Labour Council (ESLC) is a tripartite dialogue body made up of representatives from labour, business, academia and government. In January 2025, the ESLC established a special committee, the “AI and Labour Research Association,” tasked with studying changes in the labour market due to AI adoption and identifying necessary legal and policy responses. Although this discussion remains in its early stages, it is anticipated that, once a conclusion is reached, each stakeholder in the social dialogue will begin to take corresponding actions.
At company level, the Labour Standard Act stipulates that employers in Korea need the consent from workers (from the union if there is one, if not from the majority of workers) when they intend to alter the rules of employment in a way that is unfavourable to employees (i.e. leads to a deterioration in working conditions). However, there is uncertainty surrounding how AI adoption impact working conditions and what this means for the obligation of employers to consult workers.
In practice, worker consultation on AI adoption in Korea appears limited. A survey of workers in Korea showed that 56.3% were not involved in discussions around AI adoption in their workplace and only 6.9% said their firms engaged in discussions with unions or labour management councils. As expected, these discussions are far more common in large firms. In the few cases where workers were consulted, the most common topics covered were: AI’s impact on the number of jobs, changes in specific occupations, emerging training needs, and methods of data collection and use.
This section first sets out some important questions for policymakers to consider when deciding whether or not, and how, to regulate AI. It then moves on to discuss the key policy priorities and the extent to which OECD countries have tried to address challenges in these areas. Finally, the section discusses recent regulatory and policy developments in Korea and identifies policy gaps.
Regulatory considerations
Policymakers need to decide whether or not to regulate AI and, if so, whether such regulation should apply to the technology or only to certain uses of it. There are also important questions about proportionality and international spillovers to consider.
Enforcing existing regulation
The first consideration should be how to better enforce existing regulation. AI does not operate in a regulatory vacuum. Most OECD countries have laws in place that govern data protection and privacy, discrimination, employment protection, product safety, occupational safety and health, etc. In a first instance, steps should be taken to ensure that existing laws are adequately enforced. In some cases, that may mean clarifying the existing regulation and/or providing guidelines to employers and workers.
New regulation or not?
Countries will need to decide whether new regulation is required or not. The EU has been very active in this field – having adopted the EU AI Act and the Platform Work Directive. Some countries, however, are not convinced that new regulation will be required. In Japan, the government has recently passed the AI Act. Rather than introducing new regulation, this act relies on the voluntary efforts of AI developers and businesses to promote AI-related policies in Japan. In the United States, innovation has taken a front seat and the government has revoked some existing AI policies and directives that were considered barriers to American AI innovation. Other countries, like Australia, do not currently have new regulation related to AI and are considering their proposed approach to AI regulation. Canada seems to be opting for a middle path between regulation and self-governance. The proposed Artificial Intelligence and Data Act (AIDA) would set out a timeline for a new regulatory system, “In the initial years after it comes into force, the focus of AIDA would be on education, establishing guidelines, and helping firms to come into compliance through voluntary means.” This way, the government would allow time for the ecosystem to adjust to the new framework before enforcement actions are undertaken (Government of Canada, 2023[1]).
Regulating the technology, or its specific uses and consequences?
If countries decide new regulation may be necessary, then there is a choice to be made between regulating the technology itself, regardless of its applications, or regulating specific uses of the technology only (e.g. in finance or for facial recognition) or consequences (e.g. invasions of privacy or lack of transparency). Each approach has its merits and challenges, and countries have opted for different paths.
The EU AI Act is an example of AI-specific legislation. Other countries have preferred a more tailored approach. In Switzerland, for example, it is felt that “using algorithmic systems does not generally lead to entirely new challenges” and that existing regulation addresses most challenges of AI. Instead, Switzerland has opted to selectively amend certain laws (e.g. integrating AI transparency rules into existing data protection laws; modifying product liability laws) (Thouvenin et al., 2021[2]). Under the previous government, the United Kingdom acknowledged the existence of AI risks but, like Switzerland, did not feel the need for AI-specific legislation. Instead of targeting a specific technology, the United Kingdom preferred to focus on the context in which that technology was deployed (e.g. self-driving vehicles, or foundation models and LLMs), because “AI can have different impacts depending on how it's used” (Ponomarov, 2024[3]). At the same time, such regulation would be guided by cross-cutting principles published by the government. However, the current government indicated in its manifesto that it would introduce new regulation and has committed to ensuring that new surveillance technologies will not find their way into the workplace without consultation with workers (Booth, 2024[4]). In July 2024, the King’s Speech announced plans to establish “appropriate legislation to place requirements on those working to develop the most powerful [AI] models” as well as a Digital Information and Smart Data Bill which would be accompanied by reforms to data-related laws. In March 2025, an Artificial Intelligence (Regulation) Bill was (re‑)introduced into parliament which called for mandatory AI impact assessments and standardised compliance obligations, amongst others (Moreno, 2025[5]).
An important consideration policymakers will need to make, is whether to ban certain technologies, or at least the application of certain technologies. For example, Article 5 of the EU AI Act prohibits various AI “practices”, such as: “AI systems to infer emotions of a natural person in the areas of workplace” and “the use of biometric categorisation systems that categorise individually natural persons based on their biometric data to deduce or infer their […] trade union membership”.
The importance of proportionality
Regulatory intervention should be proportionate to the impact of the technology. This is why the EU AI Act has opted for a risk-based approach which imposes a gradual scheme of requirements and obligations depending on the level of risk posed to health, safety and fundamental rights. Canada (“high impact”1) would be following a similar approach. In New Zealand, the Ministry of Business, Innovation and Employment has said it “will take a light-touch, proportionate and risk-based approach to AI regulation" (New Zealand Ministry of Business, 2024[6]).
Proportionality can also be applied to firm size. The burden of regulation is likely to be heavier for small firms. In the United States, a proposed Algorithmic Accountability Act said it would focus on large firms only.2 Similarly, in Canada under the proposed AIDA, smaller firms would not be expected to have governance structures, policies, and procedures comparable to those of larger firms with a greater number of employees and a wider range of activities. In addition, in Canada, SMEs would receive assistance in adopting the practices needed to meet the requirements.
International spillovers
While countries may have different starting points, institutional set-ups as well as preferences, it is inevitable that legislative measures in one country will have spillover effects beyond its borders. The same way in which the General Data Protection Regulation (GDPR) resulted in non-EU companies adapting GDPR standards (and non-EU countries adapting comparable legislation), the EU AI Act is likely to have an impact on individuals and companies outside the EU, and to influence regulatory efforts in non-EU countries. Already, several countries are taking similar risk-based approaches to AI regulation, and legislative proposals on the table in a number of countries (e.g. Brazil, Chile) mimic elements of the EU AI Act. This impact will not just be felt where the technology is deployed, but also where it is developed. For example, regulation in the United States around discrimination in hiring may lead to developers in other countries adopting the four-fifths rule3 when designing AI hiring tools. There is a clear first-mover advantage in setting regulation, but consensus on the challenges to be addressed and the possible solutions can be promoted through international cooperation. For example, while Switzerland has chosen to rely on sectoral rules rather than AI-specific legislation, it has acknowledged that this may cause tension with international regulatory frameworks (FDFA, 2022[7]) and Switzerland is taking an active role in shaping global AI regulations, e.g. as part of the Council of Europe Committee on AI. The council of Europe’s Framework Convention on AI has been signed by a number of countries, including the United Kingdom and the United States.
The OECD of course also plays an important role in these international discussions on how to address the challenges of AI and seize its benefits. In 2019, the OECD Ministerial Council adopted the first intergovernmental standard on AI – the AI Recommendation – which aims to foster innovation and trust in AI by promoting the responsible stewardship of trustworthy AI while ensuring respect for human rights and democratic values. At the 2024 OECD Ministerial Council Meeting, ministers called on the OECD to develop an action plan to seize the benefits and address the risks of AI in the labour market.
Supporting workers through the transition
A key priority for countries will be to support workers through the transition. Like previous technological change, AI is likely to result in a transformation of the labour market (e.g. the kinds of jobs that are created, the skills that are demanded, the content of jobs and how we carry them out) and workers will need to be equipped with the tools to make the most of this transition. Most OECD countries already have strong institutions and systems in place for training, social protection, re-employment services, and social dialogue. However, AI brings new challenges which may require either a scaling-up of existing efforts, or more AI-specific solutions. This section provides an overview of the kinds of initiatives OECD countries have been undertaking, as well as the outstanding gaps. With an eye on inclusiveness (see next section) it is important also that these measures ensure no one is left behind.
Training
Employers already provide training to workers who work with AI. In a survey of workers in the manufacturing and finance sectors of seven OECD countries, nearly 50% said their employer had provided training to work with AI (Figure 3.1). At the same time, employers in those same sectors and countries identified a lack of skills as a major barrier to AI adoption, suggesting that there is scope for government intervention in this field (Figure 3.2).
Percentage of workers
Note: Workers in manufacturing and finance sectors of Austria, Germany, France, the United Kingdom, Ireland, the United States and Canada.
Source: Lane, Williams and Broecke (2023[8]), “The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers”, https://doi.org/10.1787/ea0a0fe1-en.
Share of employers
Note: Employers in manufacturing and finance sectors of Austria, Germany, France, the United Kingdom, Ireland, the United States and Canada.
Source: Lane, Williams and Broecke (2023[8]), “The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers”, https://doi.org/10.1787/ea0a0fe1-en.
Most countries recognise the importance of skills and training to make a success of AI. A recent OECD survey showed that several countries have developed AI strategies that place emphasis on skills (OECD, 2024[9]). In the United States, Executive Order 14110 “on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” (now revoked) had promised to support programmes to provide Americans with the skills they need for the age of AI. Similarly, in Norway, the National AI Strategy places emphasis on expanding the offer of education programmes and workplace training in the field of AI in order to create a solid basis of digital skills and capabilities.
Few countries, however, have proposed concrete action plans. Existing programmes tend to focus on digital or AI skills, but the share of workers requiring specialised AI skills represents only 0.3% of total employment (Green and Lamby, 2023[10]) and recent OECD analysis of training catalogues confirms that only between 0.3% and 5.5% of available training courses in Australia, Germany, Singapore and the United States deliver AI content (OECD, 2024[9]). Most workers who will be exposed to AI will not need such specialised AI skills. Skills demanded in occupations highly exposed to AI include: general project management, finance, administration and clerical skills. While the demand for these skills is still increasing overall, there is some indication that demand may be falling in the workplaces most exposed to AI (Green, 2024[11]). Analysis focussing on Canada reveals instead an increasing demand for social skills (Green, 2024[12]). Training systems will need to be agile to respond to such changes in skills demand.
Workers and the population more generally will also need basic AI literacy – i.e. a non-technical understanding of, and an ability to critically reflect on, AI applications. Article 4 of the EU AI Act requires providers and deployers of AI systems to ensure a sufficient level of AI literacy of their staff and other persons dealing with AI systems on their behalf, and the EU AI Office published “Questions & Answers” to explain what this means. Several countries have already introduced publicly funded training programmes for AI literacy (OECD, 2024[9]). In Austria, the initiative “Digital Everywhere” (Digital Überall) would roll out 3 500 workshops in all municipalities throughout 2024 with the aim of bolstering basic digital competencies, including AI and cybersecurity, among the general population. To ensure inclusivity across diverse demographics, the workshops are conducted at diverse venues, including youth centres and retirement homes, facilitating access for individuals from different backgrounds and age groups. Singapore launched a training initiative aimed at assisting jobseekers and employees from various professional backgrounds and skill levels in acquiring digital and AI literacy skills that equip them for the digital economy. The initiative is targeted at adults in jobs likely to be affected by AI and with low levels of skills and education.
Technological developments are one of the major forces behind the need for retraining, but they can also be part of the solution. In particular, AI has the potential to increase training participation, including among currently underrepresented groups, by lowering some of the barriers to training that people experience, increasing motivation to train, personalising content and delivery, and offering more targeted learning pathways. Moreover, certain AI solutions for training may improve the alignment of training to labour market needs, and help in planning and delivering training policies. To realise the benefits of AI for training and ensure that it yields benefits for all, it will be necessary to address potential drawbacks in terms of changing skills requirements, inequalities in access to data, technology and infrastructure and important ethical issues. Finally, even when these drawbacks can be addressed, the introduction and expansion of AI tools for training is constrained by the supply of AI skills in the workforce and the availability of scientific evidence regarding the benefits of AI tools for training and whether they are cost-effective (Verhagen, 2021[13]).
Social protection and re‑employment services
While AI is unlikely to lead to massive technological unemployment or a jobless future, change in the labour market will be inevitable with some jobs disappearing and new ones created. This change in the labour market will require workers to adjust, with some needing to transition between jobs. In countries without adequate social protection, efforts will need to be made to ensure that such workers are adequately supported. In the United States, Executive Order 14110 “on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” (now revoked) had stated that there was a need to “assess how current or formerly operational Federal programmes designed to assist workers facing job disruptions – including unemployment insurance and programmes authorised by the Workforce Innovation and Opportunity Act (Public Law 113‑128) – could be used to respond to possible future AI-related disruptions.” Similarly, other countries will need to ensure access to employment support measures for workers most exposed to AI-related disruption.
At the same time, AI can help improve the delivery of minimum income benefits (MIB) and unemployment assistance (UA), for instance by providing information to individuals, through determining eligibility based on pre‑determined statutory criteria and identifying undue payments, to notifying individuals about their eligibility status. One of the key opportunities of using AI for these purposes is that this may improve the timeliness and take‑up of MIB and UA. However, it may also lead to systematically biased eligibility assessments or increase inequalities, amongst others (Verhagen, 2024[14]). Similarly, AI can help improve the provision of public employment services (PES) and half of PES in OECD countries are already employing AI to enhance their services, including most commonly to match jobseekers with vacancies. While AI offers many opportunities to improve PES services, there are also risks that need to be managed, including through: prioritising the transparency of AI algorithms and explainability of results, establishing governance frameworks, ensuring end-users (staff and clients) are included and supported in the development and adoption process, and committing to rigorous monitoring and evaluation to increase the positive and manage any negative impact of AI solutions (Brioscú et al., 2024[15]).
Social dialogue
Social dialogue, including collective bargaining, can play a critical role in facilitating the trustworthy use of AI in the labour market by helping to identify opportunities for the use of AI that simultaneously achieve business objectives, mitigate risks, and safeguard human rights and international labour standards. Meaningful and early engagement with social partners provides a way for workers and worker representatives to voice questions, concerns and feedback and for social partners to negotiate to find flexible and pragmatic solutions throughout any transition period. Social dialogue will also be important to ensure that workers obtain their fair share of the benefits of AI. In general, evidence shows that the outcomes of AI for workers are more positive in firms that consult workers and their representatives on the adoption of AI (Figure 3.3).
Percentage of workers who work with AI
Note: Workers in manufacturing and finance sectors of Austria, Germany, France, the United Kingdom, Ireland, the United States and Canada.
Source: Lane, Williams and Broecke (2023[8]), “The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers”, https://doi.org/10.1787/ea0a0fe1-en.
However, AI may bring some risks to social dialogue. When employers have access to more data than necessary about workers, there is a risk that information (and hence bargaining) asymmetries are exacerbated, especially when workers are not aware that they are interacting with AI, or not sufficiently informed about the outcomes of this interaction. There is a risk also that data collected or inferred through AI is used to limit workers’ right to organise (Glass, 2024[16]). In the United States, the company Whole Foods allegedly used AI-driven data analysis to create a “unionisation heat map” to predict which stores were most likely to have union organising campaigns (CLJE, 2024[17]). Similarly, Amazon has been accused of using electronic monitoring and productivity quotas as a means to retaliate against employees engaged in union organising (Oakford, Bivens and McNicholas, 2024[18]).
This is happening in a context where the number of workers who are members of unions and are covered by collective agreements has declined in most OECD countries, and the development of new forms of work and new business models, partly facilitated by AI, risks exacerbating the under-representation challenge faced by traditional social partners (OECD, 2019[19]). The lack of AI-related expertise among social partners is a major challenge to support their members in the AI transition (OECD, 2023[20]).
Regulatory initiatives have focussed on addressing the information asymmetry between employers and workers. For example, the EU AI Act expects all employers to inform workers’ representatives and the affected workers if they will be subject to the use of the high-risk AI system. And the EU Platform Work Directive, in an attempt to regulate algorithmic management, expects information to “be given in due time to enable platform workers’ representatives to prepare for consultation, with the assistance of an expert chosen by the platform workers or their representatives in a concerted manner where needed.” In 2021, Germany amended its Works Constitution Act to give works councils a right to information if AI is used in the workplace.
Other measures that governments could implement to strengthen social dialogue include: developing guidance, tools and practical mechanisms to support the effective implementation of existing relevant regulations and establish best practices for employers (including SMEs) to involve, inform and consult workers and worker representatives in the adoption of AI systems in the workplace; supporting the development of AI-related expertise and skills among social partners; and engaging social partners in the delivery of training for AI and digital literacy initiatives for workers and management.
Promoting an inclusive transition
AI will not affect all groups equally. Some individuals are more likely to gain, while others are at higher risk of either losing their job or experiencing changes on the job (Lane, 2024[21]). The kinds of initiatives to accompany workers through the transition outlined in the previous section, particularly if well-targeted, will go some way in counteracting a potentially negative impact of AI on inclusiveness. However, there are other channels through which AI could either harm or benefit inclusiveness in the labour market, including through: the employability of underrepresented groups; bias and discrimination; and labour market concentration.
AI tools to boost the employability of underrepresented groups
Certain AI applications (e.g. live captioning for deaf individuals, AI-powered prosthetics) might help underrepresented groups achieve better outcomes on the labour market. Policies that promote the development and adoption of such tools can therefore promote inclusiveness, however these are currently lacking in OECD countries (Touzet, 2023[22]). In the case of disability, for example, Klaus Höckner from the Austrian Association of Blind and Visually Impaired People explains that “[AI] is sometimes talked about in terms of avoiding the bias, but very rarely in terms of seizing the potential” (Touzet, 2023[22]). The EU AI Act calls on member countries to support and promote research and development of AI solutions in support of socially beneficial outcomes, such as AI-based solutions to increase accessibility for persons with disabilities and tackle socio‑economic inequalities. Some policy options for governments include: funding research; help developers create sustainable business models; invest in publicly available data for the development of AI tools; promote awareness of existing solutions; and promote diversity in AI development teams (Touzet, 2023[22]).
Bias and discrimination
Existing anti-discrimination legislation is applicable to AI use in the workplace. There may, however, be gaps and loopholes in this legislation and scholars have argued that the law may need to evolve (Adams‐Prassl, Binns and Kelly‐Lyth, 2022[23]). While relevant case law is still limited, automated decision making systems are increasingly facing challenges (Adams‐Prassl, Binns and Kelly‐Lyth, 2022[23]). In the United States, for example, the Equal Employment Opportunity Commission sued and settled with iTutorGroup after it was found that the company’s recruitment software automatically rejected female applicants aged 55 and over and male applicants aged 60 and older. The bias was discovered when one of the applicants reapplied for the same job using a more recent birthdate and did get an interview (Wilkinson and Lukens, 2023[24]).
The policy response so far contains a mixture of: awareness raising, enforcement of existing legislation, impact assessments/audits, as well as data governance requirements, and (proposed) revisions to the legislation (e.g. by inverting the burden of proof). Fighting bias and discrimination will also depend on transparency and explainability of AI tools (discussed further down).
In the United States, some states are proposing modifications to their anti-discrimination law, including California, where there is a proposal by the California Civil Rights Council to hold employers liable for the use of AI-Based Employment Decision Tools that have a discriminatory impact against an applicant or employee based on a protected characteristic (Wilkinson and Lukens, 2023[24]). At the same time, California introduced AB 331 on 30 January 2023, which would prohibit employers from using automated decision tools in a way that contributes to algorithmic discrimination, and would also, amongst others, require them to perform impact assessments for automated decision tools in use, provide individuals with notice of the use of such tools and allow them to request an alternative process (Wilkinson and Lukens, 2023[24]). On 17 May 2024, Colorado enacted the first comprehensive AI legislation in the United States with a specific focus on discrimination. Developers and deployers must use reasonable care to avoid discrimination via AI systems that make, or are a substantial factor in making, a consequential decision “that has a material legal or similarly significant effect on the provision or denial to any consumer of, or the cost or terms of employment or an employment opportunity”.
However, implementing such laws can be tricky. One of the first attempts at addressing bias in automated employment decision tools was New York City’s Local Law 144 of 2021 which prohibits employers and employment agencies from using such tools unless they have been subject to a bias audit within one year of their use, information about the bias audit is publicly available, and certain notices have been provided to employees or job candidates. Six months after the implementation of the law, researchers found only 18 bias audits and 13 transparency notices from nearly 400 employers analysed (Wright et al., 2024[25]). One of the main reasons for this seemingly low compliance may be that the law narrowly focusses on tools that are used without any human oversight. It is, however, very easy to get round that by arguing there is a human in the loop somewhere (Weber, 2024[26]).
In Canada, while the Canadian Human Rights Act (CHRA) as well as provincial human rights legislation protect individuals against discrimination, it is felt that AI systems risk causing harm to historically marginalised communities on a large scale if not properly assessed for bias. The Artificial Intelligence and Data Act (AIDA), if passed, would address this risk by requiring firms to proactively identify, assess and mitigate the risk of bias on grounds prohibited by the CHRA prior to a high-impact system being made available for use.
The EU AI Act calls for AI systems to be “developed and used in a way that includes diverse actors and promotes equal access, gender equality and cultural diversity, while avoiding discriminatory impacts and unfair biases that are prohibited by Union or national law”. The EU AI Act places emphasis on the quality of data which should be “sufficiently representative, and to the best extent possible free of errors and complete in view of the intended purpose of the system” and that attention should be paid to “the mitigation of possible biases in the datasets” “to ensure that the high-risk AI system performs as intended and safely and it does not become a source of discrimination”. To achieve this, the EU AI Act states that “providers should, exceptionally, to the extent that it is strictly necessary for the purposes of ensuring bias detection and correction in relation to the high risk AI systems, […] be able to process also special categories of personal data”. The EU AI Act also calls for “appropriate data governance and management practices”, potentially with recourse to “third parties that offer certified compliance services including verification of data governance, data set integrity, and data training, validation and testing practices.” At the same time, and as mentioned above, the EU bans the use of certain systems such as those “intended to be used to detect the emotional state of individuals in situations related to the workplace”.
In Switzerland, there is an acknowledgement that the current legislation, which only prohibits discrimination by state actors, may need revision and there is a call for a “general equal treatment law that covers and sanctions discrimination by private parties, especially, companies, based on specific protected characteristics” and that the difficulty of proving discrimination “could be solved by reversing the burden of proof” (Thouvenin et al., 2021[2]).
Something else countries could do is promote diversity in the AI workforce (i.e. the subset of workers with skills in statistics, computer science and machine learning who could actively develop and maintain AI systems). Today, this workforce is primarily male (Green and Lamby, 2023[10]) and this is likely to impact the design of AI technologies, consciously or unconsciously, potentially resulting in bias and discrimination.
Labour market concentration
There are also concerns that AI may worsen labour market concentration, which is already substantial and pervasive in OECD economies (Araki et al., 2022[27]). This could result in lower wages and fewer employment opportunities for workers.
In the United States, before being revoked, Executive Order 14110 “on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” had placed importance on promoting competition in AI and other markets and called for actions to address “the risks arising from concentrated control of key inputs, taking steps to stop unlawful collusion and prevent dominant firms from disadvantaging competitors, and working to provide new opportunities for small firms and entrepreneurs. In particular, the Federal Trade Commission is encouraged […] to ensure fair competition in the AI marketplace and to ensure that consumers and workers are protected from harms that may be enabled by the use of AI.”
A related issue is the lower adoption of AI by SMEs compared to larger firms which is likely to accentuate inequalities in productivity and growth between firms, as well as between workers in those firms. Even when it comes to generative AI, where barriers to use by smaller firms are supposedly lower, larger firms are still more likely to use such technologies. In Korea, while SMEs cite incompatibility with what the company does and incompatibility with company culture as important barriers to generative AI use, as well as the cost of generative AI, a lack of skills within the company is the biggest barrier, suggesting policymakers can promote the use of generative AI among SMEs by investing in training and possibly subsidising AI adoption (OECD, forthcoming[28]).
Promoting a safe and trustworthy AI
Protecting worker health and safety
Most countries have regulations that set out employers’ obligations towards employees concerning their occupational safety and health. While details vary from country to country, employers usually have to assess risks, eliminate or reduce these risks with preventative and protective measures, and inform workers about the risks and train them. While in theory such regulations should also cover AI, there may be gaps. Also, while most countries have product liability regulations, they likely will need to be adapted to the use of AI systems. Moreover, labour inspectorates may lack the knowledge and/or capacity to address new risks posed by AI.
In the EU, Directive 89/391/EEC of 12 June 1989 introduced measures to encourage improvements in the safety and health of workers at work, including the obligation for employers to assess the occupational health and safety risks, and it laid down general principles of prevention that employers are to implement. Since AI systems potentially have significant impact on the safety and on the physical and mental health of platform workers, this directive would apply to employers using AI. However, the EU has since worked on more AI-specific legislation.
The EU AI Act is “essentially a product safety regulation” (Martens, 2024[29]): it aims to ensure that AI systems used in the EU are safe. AI systems which are judged to be high-risk (which includes AI systems used in employment, worker management and access to self-employment) will need to be assessed before being put on the market and throughout their lifecycle. To this end, technical and organisational measures should be taken, for example by designing and developing appropriate technical solutions to prevent or minimise harmful or otherwise undesirable behaviour. Those technical solutions may include mechanisms enabling the system to safely interrupt its operation (fail-safe plans) in the presence of certain anomalies or when operation takes place outside certain predetermined boundaries.
On top of the EU AI Act, the Platform Work Directive provides extra protections for workers on digital labour platforms in the EU. For example, the directive expects digital labour platforms to carry out an evaluation, at least every two years, of the impact of automated monitoring or decision making systems on the working conditions of platform workers, amongst others. These mandatory evaluations should indicate whether the safeguards of the systems are appropriate to address those risks and digital labour platforms should take appropriate preventive and protective measures. The digital labour platform should make their risk evaluation and the assessment of the mitigating measures available to platform workers, their representatives and the competent authorities.
In October 2024, the US Department of Labor issued “Principles and Best Practices for Developers and Employers” to promote worker well-being – although these remain very high-level and highlight the importance of worker consultation and transparency, clear governance systems, protecting labour rights, worker upskilling and responsible data use, amongst others. It is not clear to what extent these principles reflect the policies of the current administration.
In addressing AI-related risks to health and safety, countries could strengthen the capacity of labour authorities to supervise and enforce compliance with the law including through effective sanctions and sanctioning procedures for non-compliance, and through building their capacity to make use of AI in a trustworthy manner.
Data protection and privacy
The growing integration of AI in the world of work will likely result in greater collection and analysis of data on workers and job applicants. In some cases, data on workers is collected as a byproduct and is not the main purpose of the AI system. In other cases, AI systems may be specifically designed for measuring productivity or worker surveillance. Data collected may include information such as: worker movements, biometric data, as well as digital activities.
Workers may worry about their privacy when such data is personal. They might also worry that the data are used for purposes other than for which it was intended. Increased monitoring and surveillance could also lead to stress. In addition, AI may be used to infer personal data. As was pointed out in the United States Executive Order 14110 “on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” before it was revoked, “Artificial Intelligence is making it easier to extract, re-identify, link, infer, and act on sensitive information about people’s identities, locations, habits, and desires. Artificial Intelligence’s capabilities in these areas can increase the risk that personal data could be exploited and exposed.”
The protection of workers against privacy risks varies considerably across OECD countries and gaps remain even in those with the strongest protections.
In the EU, the GDPR regulates the use of personal data by AI systems: (i) requiring such AI systems to be developed with a clearly-defined purpose; (ii) allowing the use of personal data on the basis of only six legal grounds (consent, compliance with a legal obligation, performance of a contract, completion of a public interest mission, the safeguarding of vital interests, the pursuit of a legitimate interest); (iii) requiring data minimisation;4 (iv) limiting the retention period; (v) requiring transparency; and (vi) respecting the individuals data rights.5
In practice, data protection and privacy rules for workers across Europe are far from “harmonised, consistent, and comprehensive” because GPDR “leaves the issue of employee data protection to be addressed at the Member State level” (Abraha, 2022[30]). There are various reasons why it is hard to regulate workplace privacy through a general privacy law (Abraha, 2022[31]): (i) employers often rely on supposed legitimate interest for monitoring workers, making it harder to decide what is acceptable and what is not; (ii) the power imbalance implicit in the employer-employee relationship casts doubt over the meaning of consent in such settings; (iii) in the employment context, there are collective rights to information, participation and co-determination, and such rights are missing in general data protection regulations. Many EU member countries have no employment-specific provisions in their data protection laws. However, workers’ personal data can also be regulated through other laws (e.g. labour law) and/or collective agreements. Only a small number of countries, such as Malta, do not include any employment-specific provisions in either their respective data protection law or labour law. In some countries (e.g. Austria, Germany, Sweden) social partners play an important role in the introduction/use of new technologies for monitoring purposes, while this role is more limited in others (e.g. France and Finland). As a result, workers’ data protection in Europe is regulated through a mixture of national labour law and collective agreements, in addition to data protection law, and there are differences across countries in the extent to which workers’ data is protected (Abraha, 2022[31]).
Another challenge in the context of AI, is that the GDPR relies on general principles which leave room for interpretation. Such uncertainty combined with the high penalties of non-compliance, could discourage many companies (and in particular small and medium-sized ones) from investing in AI. Guidance from data protection bodies (or other competent authorities) would go a long way in addressing such uncertainty (Sartor and Lagioia, 2020[32]). Thus, the successful application of GDPR to AI will depend on what guidance data protection bodies and other competent authorities provide. Appropriate guidance diminishes the cost of legal uncertainty and directs companies to efficient and data protection-compliant solutions (Sartor and Lagioia, 2020[32]).
The EU platform work directive, by virtue of regulating algorithmic management, takes worker privacy a step further by forbidding: (i) the processing of personal data on the emotional or psychological state of the person performing platform work; (ii) the processing of any personal data in relation to private conversations, including exchanges with other persons performing platform work and their representatives; (iii) the collection of any personal data while the person performing platform work is not offering or performing platform work; (iv) processing personal data to predict the exercise of fundamental rights, including the right of association, the right of collective bargaining and action or the right to information and consultation; (v) the processing of any personal data to infer racial or ethnic origin, migration status, political opinions, religious or philosophical beliefs, disability, state of health, including chronic disease or HIV status, the emotional or psychological state, trade union membership, a person’s sex life or sexual orientation; and (vi) the processing of any biometric data of a person performing platform work to establish that person’s identity by comparing that data to stored biometric data of individuals in a database.
Awareness of these issues is growing, however, and some countries are considering taking steps. For example, in Germany, political discussions have been underway to enact a German national law on employee data protection in addition to the EU GDPR, which could also contain dedicated provisions on AI (White & Case, 2024[33]).
Gaps in privacy protections are greater in other OECD countries.
In the United States, data protection and privacy laws have been described as a “checkerboard of federal and state privacy” (Kerry, 2020[34]) as well as a “patchwork with varying applicability to the employment context” (Gaedt-Sheckter and Maxim Lamm, 2023[35]), leaving “workers in the U.S. virtually unprotected from employers using digital workforce management technologies” (Feng, 2023[36]).
At federal level in the United States, the Electronic Communications Privacy Act (ECPA) prohibits an employer from intentionally intercepting the oral, wire and electronic communications of employees, unless the monitoring is done for a legitimate business reason or the employer obtained the employee’s consent. In practice, this offers little protection for employees since both express and implied consent suffice, including the inclusion of a disclaimer in an employee handbook or electronic communications policy that explicitly provides notice to employees that the employee has no expectation of privacy in the use of the company’s communications systems and that the company maintains the right to monitor employee communications (Lee et al., 2022[37]).
While various states have passed privacy laws, most of these laws specifically exempt employee and job applicant data (Gaedt-Sheckter and Maxim Lamm, 2023[35]). California is a notable exception. Under the California Privacy Rights Act (CPRA), workers have the right to know when employers are monitoring them and for what purpose. They can access their data and request to correct or delete it. They can know if employers are profiling them or buying data about them, like social media activity. And they can opt out of employers selling their data (Feng, 2023[36]). There are similarities between the CPRA and the GDPR, but there are also important differences such as: the CPRA does not require a legal basis for the processing of personal data (and hence consent is not required to process information under the CPRA in the same way that processing is defined under the GDPR). In addition, the CPRA only applies to companies of a certain size (OneTrust DataGuidance and Newmeyer & Dillion LLP, 2022[38]).
In some cases, policymakers decide to ban some AI technologies to protect privacy. In the United States, for example, various cities have adopted bans on facial recognition technologies. Similarly, the EU AI Act prohibits AI systems that create or expand facial recognition databases through the untargeted scraping of facial images from the internet or CCTV footage.
Not all countries feel like data protection and privacy regulations need to change. In Switzerland, for example, the view is that the processing of personal data by algorithmic systems does not raise any fundamentally new questions and that it therefore seems possible (in principle) to solve the challenges for the protection of privacy and data using existing data protection law (Thouvenin et al., 2021[2]). Similarly, in New Zealand, the existing Privacy Act is deemed to cover AI tools and the Office of the Privacy Commissioner issued guidance on compliance with privacy law when using AI (New Zealand Privacy Commissioner, 2023[39]).
Human agency and oversight
As algorithms and automation permeate the workplace, there is a risk that individuals lose the ability to control their choices and decisions (i.e. their autonomy). Not only could this result in breaches of human rights and democratic values, but it could also undermine human dignity if it “instrumentalises” individuals or robs them of their sense of self-worth (Teo, 2023[40]).
A particular threat to agency and dignity comes from automated decision making and algorithmic management practices, which gained prominence in the platform economy but are very common in more traditional workplaces too (Milanez, Lemmens and Ruggiu, 2025[41]). The prevalence of algorithmic management tools ranges from 90% of firms in the United States to 79% in European countries and 40% in Japan. These practices involve the use of automated systems to replace functions that managers usually perform, such as allocating tasks, determining work schedules, giving instructions, monitoring and evaluating workers, as well as providing incentives or imposing sanctions. Algorithmic management can be intrusive and can restrict workers’ autonomy and diminish opportunities for human interaction, with potentially important consequences for working conditions and equal treatment at work, amongst others. It could also lead to a shift in management and power relations within the workplace.
While “a total prohibition of automated decision-making would not be feasible” (Lukács and Váradi, 2023[42]), policymakers across OECD countries are taking steps to regulate it by: clarifying the circumstances under which automated decision making may or may not be permitted, requiring human oversight throughout the AI lifecycle as well as transparency (see next section); requiring periodic testing and validation of the AI; and giving individuals rights to: human intervention, express their opinion, to obtain an explanation, and challenge/contest decisions taken by algorithms.
In the EU, automated decision making is regulated by the GDPR. More specifically, Article 22 of the GDPR states that “data subjects shall have the right not to be subject to a decision based solely on automated processing”. In practice, however, very few decisions in the employment context rely solely on automated processing. It is easy to argue that there is a “human in the loop” even if that only means the rubberstamping of AI decisions. There are also exceptions that would allow for decisions based solely on automated processing: (i) if the automated processing is necessary for entering into, or performance of a contract; (ii) when the decision based on automated processing is authorised by “Union or Member State law”; and (iii) if the data subject has given his or her explicit consent. That does not mean there are no requirements for the employer who uses those exceptions to justify decisions based solely on automated processing: individuals subject to such decisions would still have a right to obtain human intervention, to express their point of view, and to contest the decision.
The EU Platform Work Directive takes these regulations a step further. It ensures that a person performing platform work cannot be fired or dismissed based on a decision taken by an algorithm or an automated decision making system. Instead, platforms must ensure human oversight on important decisions that directly affect the persons performing platform work. The data subject has a right to obtain human intervention, to express his or her point of view, to obtain an explanation and to contest the decision. The Directive also states that digital labour platforms shall not use automated monitoring and decision making systems in any manner that puts undue pressure on platform workers or otherwise puts at risk the physical and mental health of platform workers. Digital labour platforms are expected to regularly monitor and evaluate the impact of individual decisions taken or supported by automated monitoring and decision making systems on working conditions. They are expected to tackle systematic shortcomings on the use of automated monitoring and decision making systems. When the outcome of the oversight activities identifies high risks of discrimination at work, or the infringement of rights of persons performing platform work, digital labour platforms should take appropriate measures to address them, including the possibility to discontinuing such systems.
The EU Platform Work Directive also takes steps to ensure that human oversight is meaningful. In particular, digital labour platforms should ensure sufficient human resources for human oversight and the persons charged with the function of overview should have the necessary competence, training and authority to exercise that function and the right to override automated decisions. They should be protected from dismissal, disciplinary measures or other adverse treatment for exercising their functions.
In Australia, the government has agreed with the recommendations made by the Privacy Act Review Report around automated decision making (Australian Government, 2023[43]), including: a requirement for privacy policies to set out the types of personal information that will be used in substantially automated decisions; the introduction of a right for individuals to request meaningful information about how substantially automated decisions are made. The first of these has already been implemented, through the passage of the Privacy and Other Legislation Amendment Act 2024, and will come into effect on 10 December 2026. These recommendations are very much in line with the EU GDPR except that they extend to “substantially automated” decisions, rather than being restricted to “solely automated decisions” because “few decisions are made without any level of human intervention, and if the proposal was restricted to solely automated decisions, entities could potentially bypass requirements by including a negligible level of human involvement” (Australian Government Attorney-General’s Department, 2022[44]).
In Canada there has been a Directive on Automated Decision-Making since 2019, however it only applies to tools used in administrative decision making. The directive requires algorithmic impact assessments prior to the production of an automated decision system, as well as the regular reviewing and updating of it. Other important aspects of the directive include: transparency (e.g. providing notice before decisions, or explanations after decisions) as well as recourse for clients wishing to challenge a decision.
In the United States, a proposal at federal level (the “No Robot Bosses Act”) had sought to prohibit certain uses of automated decision systems (ADS) by employers, require employers to disclose how and when ADS are being used, and to add protections for employees and applicants related to ADS. In particular, the act would have: (i) prohibited employers from relying solely on ADS in making employment-related decisions; (ii) required pre‑deployment and periodic testing and validation of ADS for issues like discrimination or bias before they are used in employment-related decisions; (iii) required employers to train individuals or entities on the operation of ADS; (iv) directed employers to provide independent human oversight of ADS outputs before using outputs in employment-related decisions; (v) required employers to timely disclose their use of ADS, the data inputs and outputs of these systems, and employee rights related to decisions aided by these systems; and (vi) established a Technology and Worker Protection Division at the Department of Labour to regulate the use of ADS in the workplace. Given the change in the policy environment in the United States, it seems unlikely that this act will become law.
At state level in the United States, some legislation around automated decision making has already been adopted, mainly around its use in recruitment, while further legislation is being proposed. As mentioned above, New York City Local Law 144 prohibits employers and employment agencies from using Automated Employment Decision Tools (AEDTs) unless they meet certain requirements around bias and transparency.
Transparency
The ability of workers to detect risks or harms, effectively question outcomes, and exercise their rights, hinges on their awareness of their interactions with AI systems and how those systems reach their outcomes. However, in many cases, workers are not aware they are interacting with an AI system. In the United Kingdom, for example, only 17% of adults can often or always tell when they are using AI (Harris et al., 2023[45]). In addition, workers (or their representatives) may not have access to information on how the AI systems work or which data they use, nor on what rights they have when interacting with such systems. Employers, in turn, may not be fully aware of how the AI systems they purchase from providers work, nor what their limitations or potential risks are.
Most AI principles that have been published to date underscore the importance of transparency of AI and its use. As put in the EU AI Act: “transparency means that AI systems are developed and used in a way that allows appropriate traceability and explainability, while making humans aware that they communicate or interact with an AI system, as well as duly informing deployers of the capabilities and limitations of that AI system and affected persons about their rights”.
To address concerns related to the opacity and complexity of AI systems, the EU AI Act requires transparency for high-risk AI systems both before and after they are placed on the market or put it into service. High-risk AI systems will have to be designed in a manner that enables deployers (in this case: employers) to understand how the AI system works, evaluate its functionality, and comprehend its strengths and limitations. All such systems should be accompanied by appropriate information in the form of instructions of use, including the characteristics, capabilities and limitations of performance of the AI system. This should assist employers in making the correct choice of system, as well as in the subsequent use of that system. Employers in turn need to provide information to workers and their representatives on the planned deployment of high-risk AI systems. In particular, individuals should be notified that they are interacting with an AI system (unless this is obvious from the point of view of an individual who is reasonably well-informed, observant and circumspect taking into account the circumstances and the context of use). The information provided by the employer should include the intended purpose and the type of decisions made by the AI system, as well as the individual’s rights under the EU AI Act.
The EU Platform Work Directive takes this a step further: “digital labour platforms should be subject to transparency and information obligations in relation to automated monitoring systems and automated systems which are used to take or support decisions that affect persons performing platform work, including platform workers’ working conditions, such as their access to work assignments, their earnings, their safety and health, their working time, their promotion or its equivalent and their contractual status, including the restriction, suspension or termination of their account. It should also be specified which kind of information should be provided to persons performing platform work regarding such automated systems, as well as in which form and when it should be provided. Individual platform workers should receive that information in a concise, simple and understandable form, in so far as the systems and their features directly affect them and, where applicable, their working conditions, so that they are effectively informed. They should also have the right to request comprehensive and detailed information about all relevant systems. Comprehensive and detailed information regarding such automated systems should also be provided to representatives of persons performing platform work, as well as to national competent authorities upon their request, in order to enable them to exercise their functions.”
The EU Platform Work Directive only applies to workers on digital labour platforms, a small share of workers overall in the EU. In Spain, however, the so-called Riders Act extends transparency requirements to all companies using algorithmic management, and not only to platform companies operating in the food delivery sector. The new Article 64.4 of the Workers Statute establishes that workers’ representatives must be informed of the “parameters, rules and instructions” that determine the work of the algorithm, which includes: how algorithms and artificial intelligence impact on working conditions, hiring decisions and layoffs, as well as the elaboration of workers’ profiles.
In Switzerland, there is a feeling that “persons interacting with algorithmic systems must be able to recognise that they are doing so […] and not with a human being” and there is a suggestion that the Data Protection Act would be the best piece of legislation to introduce such a requirement, given that a person’s interaction with an algorithmic system generally involves the processing of personal data (Thouvenin et al., 2021[2]). This is the approach of the EU in the GPDR.
Several States in the United States have introduced laws requiring employers to notify applicants and/or employees about their interactions with AI, but often these regulations do not encompass all conceivable AI applications and focus on the use of AI for recruitment or electronic monitoring only. The Illinois Artificial Intelligence Video Interview Act requires employers to notify applicants before the interview AI may be used to assess them and to provide applicants with information about how AI works and what characteristic(s) will be used to evaluate the applicants. New York City Local Law 144 requires certain notices to be provided to employees or job candidates before the use of Automated Employment Decision Tools (AEDTs).
One issue that may arise when pushing for more AI transparency is a potential tension with intellectual property and trade secret laws, which limit how much information can be disclosed. Within the EU, some commentators have argued that not all technical details of AI qualify for trade secret protection and that, even in those cases, there are exceptions to trade secret protection (Mylly, 2023[46]). Complicating this matter, however, is that trade secrets are treated differently in various jurisdictions (Kilic, 2024[47]).
Explainability
AI systems, particularly those using complex technologies like deep neural networks, yield outcomes that can be difficult or even impossible to explain. A lack of explainability can undermine the trust and confidence that people place in AI systems and the decisions that are informed by them. It also makes it difficult for individuals to provide informed consent to the use of such systems, or to identify and seek redress for adverse effects caused by AI systems in the workplace. A lack of trust and confidence, in turn, can cause worker resistance and hence hinder the adoption of AI systems in the workplace.
The platform economy is replete with examples of workers who faced “black box” algorithms and did not know the reasons for decisions taken by automated systems, while at the same time lacking the possibility to obtain an explanation, to question and discuss those decisions with a human being, or to contest them and seek rectification and redress.
The EU Platform Work Directive seeks to address this. It states that “persons performing platform work have the right to obtain an explanation from the digital labour platform for any decision taken or supported by an automated decision-making system without undue delay. The explanation, in oral or written form, shall be presented in a transparent and intelligible manner, using clear and plain language.” In practical terms, the digital labour platform needs to “provide persons performing platform work with access to a contact person designated by the digital labour platform to discuss and to clarify the facts, circumstances and reasons having led to the decision. Digital labour platforms shall ensure that such contact persons have the necessary competence, training and authority to exercise that function.” Persons performing platform work will also “have the right to request the digital labour platform to review the decisions”.
While the Platform Work Directive covers only platform workers, the EU AI Act has extended similar provisions to all workers who interact with AI systems. More specifically, the EI AI Act states that “affected persons should have the right to request an explanation when a decision is taken by the deployer with the output from certain high-risk systems […] which produces legal effects or similarly significantly affects him or her in a way that they consider to adversely impact their health, safety or fundamental rights. This explanation should be a clear and meaningful and should provide a basis for affected persons to exercise their rights”. In a way, this is an extension of the GDPR which already required that data subjects be provided with “meaningful information about the logic involved” in automated decision making processes. In practice, however, the Platform Work Directive provides far more clarity than both the EU AI Act and the GDPR on what a meaningful explanation entails.
In Canada, amendments to the Quebec Act that entered into force in September 2023 now regulate automated decision making based on the processing of personal information, and require disclosure of the occurrence of processing and the provision of information about the reasons and factors that have led to a decision (White & Case, 2024[33]).
Accountability
While rules to ensure safety will reduce risks, they do not eliminate those risks entirely and, when they materialise, there may be harm or damage. In such cases, it is important to determine who is liable, however this is made particularly difficult in the case of AI given the complexity, autonomy and opacity of such tools, which can make it difficult or expensive for victims to identify the liable person and prove the requirements for a successful liability claim, and which may deter them from claiming compensation altogether.
To address this, the EU was preparing Liability Rules for Artificial Intelligence which would have eased the burden of proof through the use of disclosure and rebuttable presumptions. It would have established for those seeking compensation for damage a possibility to obtain information on high-risk AI systems to be recorded/documented pursuant to the AI Act. In addition to this, the rebuttable presumptions would have given those seeking compensation for damage caused by AI systems a more reasonable burden of proof and a chance to succeed with justified liability claims. The Liability Rules for AI have, however, been abandoned.
In Switzerland, it is felt that “the norms of general liability law also apply to [algorithmic] systems” (Thouvenin et al., 2021[2]) and that “in certain sectors, strict liability rules that apply to algorithmic systems (e.g. for vehicles in the Road Traffic Act or drones in the Air Traffic Act) are already available” (Thouvenin et al., 2021[2]), but also there is a question as to whether “strict operator liability should be introduced for operators of algorithmic systems in other sectors” and whether “the Swiss Product Liability Act must be updated” (Thouvenin et al., 2021[2]).
According to the Canada AIDA, accountability “means that organisations must put in place governance mechanisms needed to ensure compliance with all legal obligations of high-impact AI systems in the context in which they will be used. This includes the proactive documentation of policies, processes, and measures implemented.”
In recent years, legislators across the OECD have made efforts to promote accountability mechanisms, such as requiring impact assessments and/or audits of AI systems to provide evidence and assurance that they are trustworthy and safe to use.
In Canada, the AIDA, if passed, would introduce obligations to establish and maintain written accountability frameworks by persons who make a general-purpose system or high-impact system available, or who manage the operations of such systems. Written accountability frameworks would include: (i) a description of the roles, responsibilities, and reporting structure for all personnel who contribute to making the system available or managing its operations; (ii) policies and procedures respecting the management of risks relating to the system and the data used by the system; and (iii) anything else required by regulation (White & Case, 2024[33]).
In the United States, New York City Local Law 144 expects employers using automated employment decisions tools to complete yearly bias audits.
The EU AI Act requires providers to carry out ex ante conformity assessments of high-risk AI tools before they are placed on the market, and to put in place post-market monitoring systems. The EU AI Act also places accountability with deployers (including employers) to use AI systems in accordance with the instructions of use. Indeed: “Whilst risks related to AI systems can result from the way such systems are designed, risks can as well stem from how such AI systems are used.”
Accountability is closely linked to transparency and the information about AI systems that permit traceability – including technical documentation which demonstrates the compliance of the AI system with relevant requirements, as well as the automatic recording of events (logs) over the duration of the lifetime of the system (EU AI Act).
Accountability is also linked to human oversight, to ensure that AI systems are used as intended and that their impacts are addressed over the system’s lifecycle. This includes making sure the system is subject to in-built operational constraints that cannot be overridden by the system itself and are responsive to the human operator, and that the natural persons to whom human oversight has been assigned have the necessary competence, training and authority to carry out that role.
Finally, skills and AI literacy are key to promoting accountability. The EU AI Act states that “AI literacy should equip providers, deployers and affected persons with the necessary notions to make informed decisions regarding AI systems”. In the United States, the Federal Government will work to ensure that all members of its workforce receive adequate training to understand the benefits, risks, and limitations of AI for their job functions (BEO).
The evolution of AI-related legislation in Korea
At present, no AI-specific legislation is in force in Korea; only regulatory frameworks scheduled to come into effect in 2026 have been established. Currently, the Personal Information Protection Act (PIPA) contains a limited number of provisions that relate to AI. All this reflects a broader tension within the Korean Government, mirroring trends observed internationally: the incentive to promote AI industry development, balanced against the need to regulate potential risks associated with AI.
Currently, Korea’s economic growth is driven primarily by semiconductor industries and automotive manufacturing, which account for 20.9% and 13.4% of national exports, respectively (Yoo, 2025[48]). However, the government recognises AI as a potential new pillar for future economic development. Consequently, there has been a marked reluctance to introduce stringent regulations on AI, leading to a deliberate regulatory ambiguity. Both the government and the National Assembly have been responsive to these concerns that excessive AI regulations could undermine competitiveness in both the domestic market, in which global firms are not subject to those regulations, and the international market, cautioned by domestic AI firms, notably in the software and platform sectors. Although there are growing calls for pre‑emptive regulatory measures to address emerging AI risks, the government’s emphasis remains on boosting economic growth, resulting in a cautious approach to AI regulation.
South Korea’s first law tangentially connected to AI was the PIPA, enacted in 2011. However, the original PIPA contained no AI-specific provisions, focussing instead on regulating the collection, use, and provision of personal data to safeguard individual privacy. At that time, AI was merely a possibility with limited relevance to everyday life or economic activity.
Following its enactment, there have been numerous amendments to PIPA, which were largely driven by the European Union – Korea Free Trade Agreement (EU-Korea FTA) signed in December 2015. With the General Data Protection Regulation (GDPR) entering into force in the EU in 2016, Korean companies seeking to maintain access to the European market were required to demonstrate compliance with GDPR-equivalent privacy standards. The domestic business community strongly advocated for regulations that matched – but did not exceed – GDPR standards (Joe et al., 2021[49]). Accordingly, the Korean Government and National Assembly focussed on aligning PIPA closely with GDPR requirements, and this shows Korea’s approach to digital and AI-related legislation.
Explicit references to AI first appeared in PIPA through an amendment adopted on 14 March 2023. This amendment introduced Article 37(2), which defines what constitutes an automated system, including AI, and specifies the rights of data subjects with respect to automated decision making. For example, a data subject can request an explanation on how AI made a decision and refuse to accept it, when such decision made by AI has a significant effect on his or her right or duty. Also, the criteria and procedures for making automated decisions and the methods of processing personal information must be disclosed to each data subject upon request (Korea Legislation Research Institute, 2023[50]). Nonetheless, it would be more accurate to interpret this amendment as a preparatory step towards comprehensive AI legislation, rather than a fully-fledged regulatory framework for AI.
The Korean Government and National Assembly have passed a draft AI Basic Act, set to be implemented in 2026, and are currently conducting hearings with firms, various groups, and individuals to finalise the AI Basic Act. The legislation has three primary objectives: alignment with international standards, facilitating AI development, and providing ethical and operational guidelines.
The first objective is to establish a legal framework equivalent to the EU AI Act, thereby ensuring that Korean firms can continue to benefit from the EU-Korea FTA in digital trade. Korea’s adequacy decision from the EU in 2021, confirming that PIPA provided GDPR-level protection (Jütten, 2024[51]), reflects the strategic importance of regulatory alignment in sustaining international trade relations. Similar to the motivation behind previous PIPA amendments – as an effort to ensure legal coherence aimed at promoting both exports and imports in the data sector, the proposed AI Basic Act focusses less on restricting AI and more on facilitating digital trade (European Commission, 2023[52]).
The second objective of the AI Basic Act addresses concerns from industry stakeholders, with focus on SMEs and start-ups, regarding current data restrictions under PIPA (Ministry of Science and ICT, 2024[53]). In its present form, PIPA imposes stringent conditions on the collection, processing, and use of personal data, complicating the acquisition of high-quality datasets essential for AI and machine learning. The current version of the AI Basic Act seeks to clarify the forms and conditions under which data may be used for AI development, thereby easing restrictions and promoting innovation. In particular, the law aims to exempt AI-specific data activities from some of the more burdensome requirements of PIPA, such as the need for explicit consent from each data subject for every new use.
The third objective is to provide guidelines addressing some of the ethical issues surrounding the use of AI, including privacy, transparency and safety. Evidence from a survey suggests that Korean workers believe their employers do not always disclose the use of AI (Box 3.1). While the guidelines of the AI Basic Act may appear regulatory, they are intended primarily to clarify what business practices are prohibited. Importantly, the guidelines will not be enforced directly by the government; instead, the private sector will be encouraged to develop and implement its own mechanisms for ensuring AI safety and reliability. Current proposals suggest that, although companies must assess the trustworthiness of AI systems in employment, promotion and production management, the methods for doing so will be determined autonomously. Research indicates that data-related regulations in Korea, including PIPA amendments to align with GDPR, have had greater regulatory effects than trade‑promoting effects (Lee et al., 2018[54]). This result, combined with domestic companies’ concern on being left behind global competition, suggests that strong and direct regulation in Korea remains unlikely in the immediate future.
In a survey of employees who use AI, 50.7% of employees agreed with the statement “My company is transparent about its use of AI”, while 39.9% disagreed and 4.5% strongly disagreed (Figure 3.4).
Percentage of employees reporting they agree their employer is being transparent about the use of AI in the workplace
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).
Integrating AI into Public Employment Services: The case of Korea’s Work24 Platform
Meanwhile, the Korean Government is actively advancing the integration of AI into a range of public services. As part of this broader strategy, the Ministry of Employment and Labour has begun deploying AI technologies to enhance the delivery of employment services – an approach already well-established among private sector platforms. One such initiative is Work24, a website operated jointly by the ministry and the Korea Employment Information Service, which offers AI-driven job matching and recruitment services to both job seekers and employers.
Previously known as WorkNet, Work24 serves as a public employment portal where companies may post job advertisements free of charge, and individuals can access job listings at no cost. Owing to its free‑to‑use model, the platform is primarily used by SMEs and public sector institutions, as opposed to large corporations. The predominant user base comprises medium- and lower-skilled job seekers targeting employment within SMEs, as well as older individuals seeking roles in the public sector (Kim et al., 2022[55]).
Work24 offers AI-based services through two primary channels:
AI is employed to optimise the matching process for both employers and job seekers. When users upload their CVs, and consent to data processing, the system analyses their profiles to identify potential job matches. It also suggests additional information commonly searched for by individuals with similar profiles. These recommendation algorithms are underpinned by extensive training datasets, incorporating historical data on job postings, user activity, and previous matches.
For employers, the AI system analyses job descriptions – factoring in criteria such as salary range, educational and skill requirements, job responsibilities, firm size, and industry – to suggest suitable candidates. This data-driven approach enhances the efficiency and relevance of the matching process on both sides of the labour market. Figure 3.5 illustrates the AI-driven matching process behind Work24.
Source: Ministry of Employment and Labor.
In addition to job matching, Work24 offers AI-enabled career counselling services. An AI chatbot provides tailored guidance on a broad spectrum of topics relevant to job search and career development. Given the considerable variation in required certifications and training across different occupations, it is challenging for human advisers to maintain up-to-date knowledge across all fields. The AI system overcomes this by retrieving and compiling real-time labour market information to deliver occupation-specific advice directly to users.
According to a 2025 announcement by the Korea Employment Information Service, the number of job seekers who secured employment through Work24 rose by nearly 31% in 2023–reaching 75 546 individuals, compared to 57 844 in 2022. This growth is attributed to the integration of AI into the platform’s core services (Korea Employment Information Service, 2025[56]).
Nevertheless, limitations remain. Work24 primarily features job opportunities within SMEs, which tend to offer lower wages than larger firms. Consequently, the platform may be less effective in connecting highly skilled workers with high-paying positions. Despite this, Work24 plays a vital role in addressing the employment needs of the SME sector, often characterised by financial constraints and under-resourced recruitment capacities (Yang et al., 2024[57]).
Social dialogue on AI in Korean workplaces
At present, the AI Basic Act is a comprehensive legislation rather than a collection of specific guidelines, particularly with regards to labour markets. Consequently, potential AI challenges in the workplace are currently expected to be addressed through the Labour Standards Act and/or PIPA.
Although there is still a chance that the AI Basic Act scheduled for implementation in 2026 incorporates labour-related provisions, it is expected that many practical issues will continue to be handled via executive orders or ordinances. Specific discussions on these matters are ongoing through the Economic, Social and Labour Council (ESLC), a tripartite dialogue body involving representatives from labour, business, academia and government. In January 2025, the ESLC established a special committee, the “AI and Labour Research Association,” tasked with studying changes in the labour market due to AI adoption and identifying necessary legal and policy responses (Economic, Social and Labor Council of Rep. of Korea, 2025[58]). Although this discussion remains in its early stages, it is anticipated that, once a conclusion is reached, each stakeholder in the social dialogue will begin to take corresponding actions.
An obligation to consult employees about unfavourable changes to the rules of employment?
A distinctive feature of labour law in Korea, as enshrined in the Labour Standards Act, pertains to provisions governing unfavourable changes to rules of employment. While many legal frameworks allow both employers and employees to negotiate the terms and conditions related to the rules of employment or working conditions, which include working hours, holidays, leave, wage determination, payment schedules, employee training, and matters concerning family allowances or retirement, the authority to unilaterally determine employment conditions rests exclusively with the employer (Article 93 of the Korean Labour Standards Act).
This inherent unilateralism creates a potential incentive for employers to modify employment rules in ways that could disadvantage employees. To mitigate this risk, labour laws in Korea mandate that when employers introduce amendments to employment rules or conditions that negatively affect employees, those employees must be recognised as key stakeholders in the decision making process. Article 94 of Korea’s LSA explicitly outlines this requirement (Procedures for Preparation and Amendment of Rules):
1. An employer shall, with regard to the preparation or alteration of the rules of employment, hear the opinion of a trade union if there is such a trade union composed of the majority of the employees in the firm or workplace concerned, or otherwise hear the opinion of the majority of the said employees if there is no trade union composed of the majority of the employees: provided, that in case of amending the rules of employment unfavourably to employees, the employer shall obtain their consent thereto.
2. When an employer reports the rules of employment pursuant to Article 93, he or she shall attach a document stating the opinion as referred to in paragraph (1).
Consequently, any business activity resulting in wage reductions, increased working hours, deteriorated working conditions, or a weakening of employees’ collective bargaining power requires the consent of either a trade union representing the majority of employees or, in the absence of such a union, the majority of employees themselves. A landmark ruling by the Korean Supreme Court (Decision 77Da355, 26 February 1977) underscored that the legitimacy of unfavourable changes is determined not by the outcome itself (i.e. benefit or harm), but by whether the employer and employees negotiated from a position of equality. The legislative intent behind this provision is to ensure that employees, who intrinsically occupy a subordinate position within the employment relationship, are guaranteed collective participation when employment conditions are modified to their detriment (Kang, 2025[59]).
The adoption of AI in the workplace presents several potential legal challenges regarding unfavourable changes to the rules of employment. The main issue concerns the uncertainty surrounding how AI adoption impacts working conditions. When an employer decides to implement AI, it may not be immediately clear whether this will worsen or improve employees’ working conditions. For instance, while AI could enhance productivity in manufacturing processes, potentially leading to layoffs – an action unequivocally requiring employee consent, AI might also result in reduced working hours and corresponding wage cuts, or unanticipated increases in work intensity across the production process.
This scenario raises critical questions: do such changes also necessitate employee consent? If so, must this consent be secured in advance, or is ex post facto consent sufficient? Furthermore, when unfavourable changes in employment conditions are implemented not through explicit amendments in contracts but through practical shifts within an otherwise unchanged set of employment rules, it remains ambiguous whether such changes should legally be regarded as a revision of employment rules. Additional questions arise, such as whether employees possess the right to refuse to work with AI, and whether compelling the use of AI constitutes an unfavourable change to the rules of employment.
While past court rulings have frequently categorised mixed changes – those entailing both favourable and unfavourable elements – as unfavourable, more recent judicial interpretations have shown a tendency to avoid requiring collective consent solely on the basis of a change being unfavourable (Kim, 2022[60]). Nevertheless, ambiguity persists for employers regarding the threshold at which changes become sufficiently significant to trigger the requirement for collective consent.
The Korean Labour Standards Act permits changes to employment conditions via rules of employment – even if unfavourable – when the possibility of such changes is explicitly stated in the employment contract (Kim, 2020[61]). However, it remains unclear whether this legal basis applies to mid- or long-term changes arising from AI adoption. This uncertainty is particularly problematic for SMEs, which lack comprehensive human resources management and legal teams. Consequently, SMEs may delay AI adoption until clearer legal guidance emerges – whether through case law, policy frameworks, or interpretations under Korea’s nascent AI Basic Act. The hesitation of AI adoption from legal uncertainty may partly explain the significantly lower AI adoption rates observed in SMEs in Korea. However, given the considerable productivity gap between SMEs and large firms, and the substantial potential benefits of AI adoption in the SME, addressing this hesitation is crucial.
A further legal issue pertains to the decision making process for unfavourable changes to employment rules, which operates on the principle of majority rule. While requiring consent from a majority of employees or a union representing the majority promotes democratic decision making (Lee, 2022[62]), it also creates a potential for decisions that disproportionately disadvantage minority groups. For example, a majority union may consent to employment rule changes that impose burdens on non-union employees while concentrating benefits among union members (Kang, 2023[63]). In the context of AI adoption, a union representing the majority might request that AI be deployed preferentially (or last) in production lines or workplaces where non-union members are concentrated. Such outcomes would not align with the original intent of Article 94 of Labour Standards Act and may, in the future, raise concerns regarding fairness and inclusiveness in employment governance.
AI education and training policies in Korea
To address the growing demand for skilled labour in AI and digital technologies, the Korea Ministry of Employment and Labour has implemented a range of education and training programmes with targeted public support through various institutions (Box 3.2). However, three challenges going forward will be: (i) to achieve better strategic co‑ordination between the Ministry of Trade, Industry and Energy and the Ministry of Employment and Labour; (ii) to provide training programmes tailored to the specific needs of SMEs; and (iii) to promote more work-based learning.
Education programmes
“New Technologies Education for High-school Students” offers training opportunities to tenth-grade students enrolled in high schools that have restructured their curricula for new technologies, including AI. The programme provides financial support for students to take courses related to new technologies, as well as for teachers to participate in training programmes necessary to deliver such courses at schools. In 2024, of the total 3 138 participating students, 324 completed coursework specifically related to AI.
“High-tech Courses in Polytechnics” subsidises the costs of delivering advanced digital training courses to individuals under the age of 39 enrolled in polytechnic institutions that have established high-tech AI departments. In 2024, 215 young jobseekers completed AI-related training through this programme.
The government provides financial assistance to polytechnics offering two‑year degree programmes focussed on applied AI skills. This initiative “Two-Year AI Degree Programme in Polytechnics” aims to rapidly cultivate a workforce equipped with practical competencies. As of 2024, a provisional total of 206 students were enrolled in these AI degree tracks. The concentration of AI education in polytechnic institutions reflects the urgency of mitigating labour shortages in the AI sector through accelerated training pathways.
Training programmes
The Ministry of Employment and Labour is implementing demand-responsive, project-based training initiatives in co‑operation with leading companies, universities, and private institutions. These efforts fall under the umbrella of the “K-Digital Training programme”, which is open to all individuals who are eligible for the National Training Card system1 – both job seekers and incumbent workers.
Training courses are categorised by skill level, including basic, employed, and advanced programmes, with tuition subsidies of up to USD 250 per month. In 2024, over 5 000 individuals enrolled in 211 AI-related courses through this scheme.
For individuals requiring foundational digital skills, the “K-Digital Beginner-level Skill Training” programme is available under the same eligibility criteria. It offers additional tuition support of up to USD 400 annually for introductory-level courses. In 2024, more than 16 000 participants received AI training across 64 entry-level digital courses.
To better align workforce skills with regional industrial needs, the “Region-Industry Matching Training Programme” targets incumbent workers by offering customised training based on regional demand assessments conducted by local training centres. In 2024, of the total 54 494 participants, approximately 1 400 individuals received AI-specific training.
1. The National Training Card (NTC) system provides financial assistance ranging from USD 2 000 to USD 3 500 over a period of up to five years to individuals in need of vocational education and training. The level of support varies between 45% and 100% of training costs, determined by factors such as household income, age, labour market status, and eligibility for the Earned Income Tax Credit (EITC). Extended coverage is available for individuals from low-income households, EITC-eligible recipients, and those aged under 35 or over 60. Prior to 2020, separate systems were provided to the employed and unemployed, and the system excluded certain groups such as short-time workers, dependent self-employed persons, and small business owners. Following reforms, all eligible applicants now receive standardised support regardless of employment status. However, the NTC remains unavailable to specific categories, including: public officials, employees of private schools, students below upper secondary level (pre‑grade 12), first- and second-year university students, self-employed individuals with annual incomes exceeding USD 300 000, employees aged 45 or under working in large enterprises (with 1 000 or more employees) earning USD 2 000 or more per month, and independent contractors with monthly incomes exceeding USD 3 500.
Source: Ministry of Employment and Labour of Korea.
Strategic co‑ordination amongst ministries
Industrial policies related to AI are primarily designed and implemented by the Ministry of Science and ICT and the Ministry of Trade, Industry and Energy, while human resource policies fall under the remit of the Ministry of Employment and Labour. This structural separation can lead to a lack of alignment between industrial policy and human resource supply strategies, potentially diminishing the effectiveness of both.
Korea already faces a chronic shortage of skilled AI talent, exacerbated by brain drain, which is undermining the competitiveness of several leading AI companies (Song and Song, 2015[64]). The analysis conducted in this report, utilising Korean employment insurance data, also revealed the potential for labour demand to outstrip supply for generative AI roles. Although the rapid emergence of generative AI means that labour demand is not yet a major concern, it underscores that the supply of labour for advanced AI technologies remains insufficient. Hence, beyond simply increasing the number of AI graduates, there is a need for more strategic co‑ordination.
To address this issue, a specialised, overarching AI agency should be established to co‑ordinate the policies of various ministries. Without such co‑ordination, ministries may implement redundant or conflicting policies, reducing the effectiveness of AI-related initiatives. A dedicated AI agency could maximise policy outcomes by linking industrial policies across ministries and enhancing the effectiveness of AI education and training programmes through a more cohesive approach to human resource and industrial policies. Moreover, as AI raises both ethical and technical challenges, the involvement of an independent, specialised agency could address these issues more effectively than individual ministries.
Delivering training programmes tailored to the needs of SMEs
Han (2023[65]) indicates a strong demand for AI education, particularly its practical application, suggesting a need for AI training that is specifically tailored to the distinct requirements of individual companies. Large corporations, with a significant number of employees and substantial AI utilisation, can afford to design and offer AI education and training that addresses their employees’ needs. However, SMEs often lack the financial resources and capacity to provide the necessary AI programmes independently. Disparities in training and education opportunities between large corporations and SMEs exacerbate the productivity gap between them (Van de Wiele, 2010[66]).
In Korea, employees can enrol in AI-related education through private or government-operated educational institutions via the National Training Card system. However, due to the group-based nature of these educational programmes, most institutions tend to focus on delivering general AI knowledge rather than providing training tailored to specific company needs. Consequently, according to data provided by the Ministry of Employment and Labour on AI-related training programmes, less than 14% of 37 566 participants in the K-Digital Training programme and less than a quarter of 95 532 participants in the K-Digital Basic Competency Training programme undertook AI-related courses.
The Ministry of Employment and Labor operates Regional Employment Agencies in each province and Employment Centres at the city level to provide employment services and administration. Therefore, by gathering feedback on training programmes and collecting demand for training services from job seekers and firms at these Regional Employment Agencies or Employment Centres, it is possible to identify educational and training needs at the local level. Granting autonomy to these local units to design training programmes based on this information would enable the creation of region-specific, customised training programmes.
Therefore, government-supported institutions and publicly funded educational entities should design and offer training programmes that reflect the specific needs of SMEs within their respective regions. By aligning the content of training programmes with the industrial characteristics of local SMEs, the Ministry of Employment and Labour can increase the incentive for SME employees to participate, thereby enhancing programme effectiveness. Localised and customised training will boost the productivity of SME workers, narrow the productivity gap between SMEs and larger corporations, and ultimately contribute to the productivity of the entire economy (O’Regan, Stainer and Sims, 2010[67]; Antonioli, Mazzanti and Pini, 2010[68]). This approach will also increase the adoption of AI within SMEs and improve the utilisation of AI within those SMEs that have already implemented it.
Transitioning from AI education in institutions to on-the‑job training
Contemporary AI tools such as ChatGPT and Google Gemini afford users a high degree of operational flexibility. The scope and methods of AI usage vary significantly depending on the task, role, company, and industry. Consequently, while generic AI education has limited applicability in firms, there is growing demand for specialised, job-specific training (Maity, 2019[69]). The utilisation of AI technologies is also linked to the accumulation of firm-specific human skills (Czarnitzki, Fernández and Rammer, 2023[70]), thereby increasing firms’ demand for firm-specific training (Kourad, 2024[71]). Moreover, even for general skills, employees’ productivity can vary significantly depending on how these skills are applied in specific business contexts, leading companies to place a premium on on-the‑job training (Lazear, 2009[72]).
The Ministry of Employment and Labour’s K-Digital Training Programme already facilitates project-based learning in collaboration with firms. Similarly, training courses operated in conjunction with polytechnic institutions, such as high-tech courses in polytechnics and two‑year AI degree programmes in polytechnics, have moved towards practice‑based instruction. However, many training courses continue to focus on theoretical or general AI knowledge.
To address this, the government should shift its support from institutionalised training towards work-based learning. Such a transition would help reduce mismatches between labour supply and demand. Firms could train potential hires in firm-specific skills at lower cost, supported by public funding. In turn, this would increase firms’ incentives to hire trainees and improve post-training employment outcomes. For job seekers, particularly young people, this approach offers valuable exposure to employers’ actual needs and increases job-match quality, potentially reducing early-career turnover (Lynch, 1991[73]) and increasing wage levels (Langer and Wiederhold, 2023[74]).
By operating on-the‑job training programmes and gathering relevant data, the government can better understand the skills and expertise in demand within the private sector. This information can inform the development of rational human resource supply strategies and educational policies, improving the alignment of industrial policies with the needs of the labour market.
References
[31] Abraha, H. (2022), “A pragmatic compromise? The role of Article 88 GDPR in upholding privacy in the workplace”, International Data Privacy Law, Vol. 12/4, pp. 276-296, https://doi.org/10.1093/idpl/ipac015.
[30] Abraha, H. (2022), The GDPR and Privacy in the Workplace, https://blogs.law.ox.ac.uk/oblb/blog-post/2022/09/gdpr-and-privacy-workplace (accessed on 31 May 2024).
[23] Adams‐Prassl, J., R. Binns and A. Kelly‐Lyth (2022), “Directly Discriminatory Algorithms”, The Modern Law Review, Vol. 86/1, pp. 144-175, https://doi.org/10.1111/1468-2230.12759.
[68] Antonioli, D., M. Mazzanti and P. Pini (2010), “Productivity, innovation strategies and industrial relations in SMEs. Empirical evidence for a local production system in northern Italy”, International Review of Applied Economics, Vol. 24/4, pp. 453-482, https://doi.org/10.1080/02692171.2010.483790.
[27] Araki, S. et al. (2022), “Monopsony and concentration in the labour market”, in OECD Employment Outlook 2022: Building Back More Inclusive Labour Markets, OECD Publishing, Paris, https://doi.org/10.1787/0ecab874-en.
[43] Australian Government (2023), Government response to the Privacy Act Review Report, https://www.ag.gov.au/sites/default/files/2023-09/government-response-privacy-act-review-report.PDF (accessed on 11 June 2024).
[44] Australian Government Attorney-General’s Department (2022), Privacy Act Review Report 2022, https://www.ag.gov.au/sites/default/files/2023-02/privacy-act-review-report_0.pdf (accessed on 11 June 2024).
[4] Booth, H. (2024), “What We Know About the New U.K. Government’s Approach to AI”, Time Magazine, https://time.com/6997876/uk-labour-ai-kyle-starmer/ (accessed on 16 August 2024).
[15] Brioscú, A. et al. (2024), “A new dawn for public employment services: Service delivery in the age of artificial intelligence”, OECD Artificial Intelligence Papers, No. 19, OECD Publishing, Paris, https://doi.org/10.1787/5dc3eb8e-en.
[17] CLJE (2024), Worker Power and Voice in the AI Response, Center for Labor and a Just Economy at Harvard Law School, https://clje.law.harvard.edu/app/uploads/2024/01/Worker-Power-and-the-Voice-in-the-AI-Response-Report.pdf (accessed on 8 August 2024).
[70] Czarnitzki, D., G. Fernández and C. Rammer (2023), “Artificial intelligence and firm-level productivity”, Journal of Economic Behavior and Organization, Vol. 211, pp. 188-205, https://doi.org/10.1016/j.jebo.2023.05.008.
[58] Economic, Social and Labor Council of Rep. of Korea (2025), Social Dialogue Brief, https://english.eslc.go.kr/bbs/data/view.do?menu_idx=2100&bbs_mst_idx=BM0000000357&data_idx=BD0000000062 (accessed on 7 May 2025).
[52] European Commission (2023), Staff working document on digital trade negotiations with the Republic of Korea and with Singapore, https://eur-lex.europa.eu/legal-content/IT/TXT/?uri=CELEX%3A52023SC0085 (accessed on 5 May 2025).
[7] FDFA (2022), Artificial intelligence and international rules - Report for the Federal Council, Federal Department of Foreign Affairs, https://www.newsd.admin.ch/newsd/message/attachments/71099.pdf (accessed on 21 August 2024).
[36] Feng, K. (2023), Overview of New Rights for Workers under the California Consumer Privacy Act, https://laborcenter.berkeley.edu/overview-of-new-rights-for-workers-under-the-california-consumer-privacy-act/ (accessed on 7 June 2024).
[35] Gaedt-Sheckter, C. and E. Maxim Lamm (2023), AI in Employment: Privacy Regulation Is Here, https://www.gibsondunn.com/wp-content/uploads/2023/09/Gaedt-Sheckter_PLI-Chronicle_August_2023.pdf (accessed on 5 June 2024).
[16] Glass, A. (2024), Unions Give Workers a Voice Over How AI Affects Their Jobs, https://www.americanprogress.org/article/unions-give-workers-a-voice-over-how-ai-affects-their-jobs/ (accessed on 19 August 2024).
[1] Government of Canada (2023), The Artificial Intelligence and Data Act (AIDA) – Companion document, https://ised-isde.canada.ca/site/innovation-better-canada/en/artificial-intelligence-and-data-act-aida-companion-document (accessed on 8 August 2024).
[12] Green, A. (2024), “Artificial intelligence and the changing demand for skills in Canada: The increasing importance of social skills”, OECD Artificial Intelligence Papers, No. 17, OECD Publishing, Paris, https://doi.org/10.1787/1b20cdb6-en.
[11] 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.
[10] 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.
[65] 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.
[45] Harris, B. et al. (2023), Public awareness, opinions and expectations about artificial intelligence: July to October 2023, Office for National Statistics, https://www.ons.gov.uk/businessindustryandtrade/itandinternetindustry/articles/publicawarenessopinionsandexpectationsaboutartificialintelligence/julytooctober2023 (accessed on 12 June 2024).
[49] Joe, D. et al. (2021), “Ten Years of Korea-EU FTA: Achievements and Way Forward”, SSRN Electronic Journal, https://doi.org/10.2139/SSRN.4270056.
[51] Jütten, M. (2024), The EU’s digital trade policy, European Parliamentary Research Service.
[59] Kang, J. (2025), “Changes in the Rulings on Disadvantageous Changes to Employment Rules – Focusing on a Comparative Review with Japanese Law –”, The Journal of Labor Law, Vol. 63, pp. 115-155.
[63] Kang, M. (2023), Labor Movement of Youth Generation and Workplace Democracy, Korea Labor Institute.
[34] Kerry, C. (2020), Protecting privacy in an AI-driven world, Brookings Research, https://www.brookings.edu/articles/protecting-privacy-in-an-ai-driven-world/ (accessed on 5 June 2024).
[47] Kilic, B. (2024), “Into Uncharted Waters: Trade Secrets Law in the AI Era”, CIGI Papers, No. 295, Centre for International Governance Innovation, https://www.cigionline.org/publications/into-uncharted-waters-trade-secrets-law-in-the-ai-era/ (accessed on 13 June 2024).
[60] Kim, D. (2022), “Time Flexibility and Change the Disadvantage of Employment Rules”, Wonkwang Law, Vol. 38/2, pp. 129-150, https://doi.org/10.22397/wlri.
[55] Kim, D. et al. (2022), 2022 KEIS- Employment DB Analysis, Korea Employment Information Service.
[61] Kim, K. (2020), “Die Änderung der Arbeitsordnung mit der Ungünstigkeit und der Arbeitsvertrag”, Labor Law Forum, Vol. 29/29, pp. 21-41, https://doi.org/10.46329/LLF.2020.02.29.21.
[56] Korea Employment Information Service (2025), AI-based Job Matching Service, https://www.keis.or.kr/keis/en/bbs/300/detail.do?pstSn=5&pageIndex=1&pageItm=10&searchOrderSort=0&searchGbn=0 (accessed on 27 June 2025).
[50] Korea Legislation Research Institute (2023), 개인정보 보호법 (PERSONAL INFORMATION PROTECTION ACT), https://elaw.klri.re.kr/kor_service/lawView.do?hseq=62389&lang=KOR (accessed on 27 June 2025).
[71] Kourad, H. (2024), “Technologically intensive SMEs and AI: determinants of the growing demand for specialized training profiles”, International Journal of Strategic Management and Economic Studies (IJSMES), Vol. 3/3, https://doi.org/10.5281/zenodo.14034447.
[21] 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.
[8] 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.
[74] Langer, C. and S. Wiederhold (2023), “The Value of Early-Career Skills”, CESifo Working Paper, No. 10288, https://doi.org/10.2139/SSRN.4369987.
[72] Lazear, E. (2009), “Firm-specific human capital: A skill-weights approach”, Journal of Political Economy, Vol. 117/5, pp. 914-940, https://doi.org/10.1086/648671.
[62] Lee, D. (2022), “The requirement for changing the disadvantage of the employment rules and the favorable condition-first principle”, The Journal of Labor Law, Vol. 55, pp. 143-180.
[37] Lee, J. et al. (2022), A Privacy and Employment Law Primer: Recent Updates on Discrimination and Privacy Implications of Technology in the Workplace, Loeb & Loeb, https://www.loeb.com/en/insights/publications/2022/08/recent-updates-on-discrimination-and-privacy-implications-of-technology-in-the-workplace (accessed on 5 June 2024).
[54] Lee, K. et al. (2018), “Cross-Border Data Flows: Discussions and Countermeasures”, Korea Institute for International Economic Policy, https://doi.org/10.2139/SSRN.3422585.
[42] Lukács, A. and S. Váradi (2023), “GDPR-compliant AI-based automated decision-making in the world of work”, Computer Law & Security Review, Vol. 50, p. 105848, https://doi.org/10.1016/j.clsr.2023.105848.
[73] Lynch, L. (1991), “The Role of Off-the-Job vs. On-the-Job Training for the Mobility of Women Workers”, The American Economic Review, Vol. 81/2, pp. 151-156, http://www.jstor.org/stable/2006844.
[69] Maity, S. (2019), “Identifying Opportunities for Artificial Intelligence in the Evolution of Training and Development Practices”, Journal of Management Development, Vol. 38/8.
[29] Martens, B. (2024), The European Union AI Act: premature or precocious regulation?, Bruegel, https://www.bruegel.org/analysis/european-union-ai-act-premature-or-precocious-regulation (accessed on 30 May 2024).
[41] 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.
[53] Ministry of Science and ICT (2024), Press Releases - A New Chapter in the Age of AI: Basic Act on AI Passed at the National Assembly‘s Plenary Session, https://www.msit.go.kr/eng/bbs/view.do?sCode=eng&mId=4&mPid=2&pageIndex=&bbsSeqNo=42&nttSeqNo=1071&searchOpt=ALL&searchTxt= (accessed on 27 June 2025).
[5] Moreno, N. (2025), The Artificial Intelligence (Regulation) Bill: Closing the UK’s AI Regulation Gap?, https://kennedyslaw.com/en/thought-leadership/article/2025/the-artificial-intelligence-regulation-bill-closing-the-uks-ai-regulation-gap/ (accessed on 15 May 2025).
[46] Mylly, U. (2023), “Transparent AI? Navigating Between Rules on Trade Secrets and Access to Information”, IIC - International Review of Intellectual Property and Competition Law, Vol. 54/7, pp. 1013-1043, https://doi.org/10.1007/s40319-023-01328-5.
[6] New Zealand Ministry of Business, I. (2024), Approach to work on Artificial Intelligence, https://www.mbie.govt.nz/dmsdocument/28913-approach-to-work-on-artificial-intelligence-proactiverelease-pdf (accessed on 7 January 2025).
[39] New Zealand Privacy Commissioner (2023), Artificial intelligence and the Information Privacy Principles, https://www.privacy.org.nz/assets/New-order/Resources-/Publications/Guidance-resources/AI-Guidance-Resources-/AI-and-the-Information-Privacy-Principles.pdf (accessed on 16 January 2025).
[18] Oakford, P., J. Bivens and C. McNicholas (2024), Federal AI legislation: An evaluation of existing proposals and a road map forward, Economic Policy Institute, https://doi.org/10.1111/IREL.12286.
[9] 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.
[20] OECD (2023), OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market, OECD Publishing, Paris, https://doi.org/10.1787/08785bba-en.
[19] OECD (2019), OECD Employment Outlook 2019: The Future of Work, OECD Publishing, Paris, https://doi.org/10.1787/9ee00155-en.
[28] OECD (forthcoming), Generative AI and the SME Workforce: New Survey Evidence, OECD Publishing, Paris.
[38] OneTrust DataGuidance and Newmeyer & Dillion LLP (2022), Comparing privacy laws: GDPR v. CCPA & CPRA, https://www.dataguidance.com/sites/default/files/gdpr_v_ccpa_and_cpra_v6.pdf (accessed on 7 June 2024).
[67] O’Regan, N., L. Stainer and M. Sims (2010), “Training in SMEs and its relationship to profitability”, International Journal of Human Resources Development and Management, Vol. 10/2, pp. 166-181, https://doi.org/10.1504/IJHRDM.2010.031442.
[3] Ponomarov, K. (2024), How to Navigate UK AI Regulations, Legal Nodes, https://legalnodes.com/article/uk-ai-regulations (accessed on 20 June 2024).
[32] Sartor, G. and F. Lagioia (2020), The impact of the General Data Protection Regulation (GDPR) on artificial intelligence, European Union, Brussels, https://www.europarl.europa.eu/RegData/etudes/STUD/2020/641530/EPRS_STU(2020)641530_EN.pdf (accessed on 31 May 2024).
[64] Song, H. and E. Song (2015), “Why Do South Korea’s Scientists and Engineers Delay Returning Home? Renewed Brain Drain in the New Millennium”, Science, Technology and Society, Vol. 20/3, pp. 349-368, https://doi.org/10.1177/0971721815597140.
[40] Teo, S. (2023), “Human dignity and AI: mapping the contours and utility of human dignity in addressing challenges presented by AI”, Law, Innovation and Technology, Vol. 15/1, pp. 241-279, https://doi.org/10.1080/17579961.2023.2184132.
[2] Thouvenin, F. et al. (2021), A Legal Framework for Artificial Intelligence, University of Zurich, https://www.dsi.uzh.ch/en/research/projects/strategy-lab/strategy-lab-21.html (accessed on 30 May 2024).
[22] 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.
[66] Van de Wiele, P. (2010), “The impact of training participation and training costs on firm productivity in Belgium”, International Journal of Human Resource Management, Vol. 21/4, pp. 582-599, https://doi.org/10.1080/09585191003612083.
[14] Verhagen, A. (2024), “Using AI to manage minimum income benefits and unemployment assistance: Opportunities, risks and possible policy directions”, OECD Artificial Intelligence Papers, No. 21, OECD Publishing, Paris, https://doi.org/10.1787/718c93a1-en.
[13] Verhagen, A. (2021), “Opportunities and drawbacks of using artificial intelligence for training”, OECD Social, Employment and Migration Working Papers, No. 266, OECD Publishing, Paris, https://doi.org/10.1787/22729bd6-en.
[26] Weber, L. (2024), “New York City Passed an AI Hiring Law. So Far, Few Companies Are Following It.”, The Wall Street Journal, https://www.wsj.com/business/new-york-city-passed-an-ai-hiring-law-so-far-few-companies-are-following-it-7e31a5b7 (accessed on 30 May 2024).
[33] White & Case (2024), AI Watch: Global regulatory tracker, https://www.whitecase.com/insight-our-thinking/ai-watch-global-regulatory-tracker (accessed on 16 August 2024).
[24] Wilkinson, C. and D. Lukens (2023), The Growing Regulation of AI-Based Employment Decision Tools, Perkins Coie News & Insights, https://www.perkinscoie.com/en/news-insights/the-growing-regulation-of-ai-based-employment-decision-tools.html (accessed on 30 May 2024).
[25] Wright, L. et al. (2024), “Null Compliance: NYC Local Law 144 and the challenges of algorithm accountability”, The 2024 ACM Conference on Fairness, Accountability, and Transparency, https://doi.org/10.1145/3630106.3658998.
[57] Yang, J. et al. (2024), A Study on the Development Strategies for AI-based Digital Employment Services, Korea Employment Information Service.
[48] Yoo, S. (2025), “Contribution of Export to National Economy”, Trade Focus, Vol. 6.
← 1. While the Canadian approach to identifying “high-impact systems” would be similar to the EU’s, the categories would be slightly different. An AI system will be considered “high-impact” if it is used: (i) to determine employment matters (i.e. to determine employment, recruitment, referral, hiring, remuneration, promotion, training, apprenticeship, transfer or termination); (ii) in matters relating to service access; (iii) to process biometric information in matters relating to the identification of an individual or the assessment of an individual’s behaviour or state of mind; (iv) in matters relating to content moderation and prioritisation; (v) in healthcare or emergency services matters, with certain exceptions, pursuant to the Food and Drugs Act; (vi) by a court or administrative body in determinations relating to an individual who is a party to proceedings; and (vii) to assist a peace officer, as defined in the Criminal Code, in the exercise and performance of their law enforcement powers, duties, and functions (White & Case, 2024[33]).
← 2. For example, in the case of automated decision systems, the law would apply only to persons, partnerships, or corporations with greater than USD 5 000 000 in average annual gross receipts or greater than USD 25 000 000 in equity value for the 3‑taxable‑year period preceding the most recent fiscal year.
← 3. In the United States, the four‑fifths rule is a guideline used to determine if there is adverse impact in the selection process of a specific group. The rule states that the selection ratio of a minority group should be at least four‑fifths (80%) of the selection ratio of the majority group.
← 4. I.e. The personal data collected and used must be appropriate, relevant and limited to what is necessary for the defined objective.
← 5. These rights include: access, rectification, erasure, restriction, portability and objection.