Occupational exposure to artificial intelligence and worker preparedness: Skills, educational attainment, and adult learning across 16 Countries

Abstract:

Debates about artificial intelligence (AI) in the workplace have largely focused on identifying which occupations are most exposed, while paying less attention to whether workers currently employed in those occupations are adequately prepared to adapt. This study links data on 52,032 employed adults aged 25 - 65 across 16 countries from the OECD 2023 Survey of Adult Skills with the OECD 2026 occupational AI exposure measure. Contrary to the prespecified hypotheses, greater occupational exposure to AI was associated with lower adaptive problem-solving proficiency (-11.89 points for each 0.10 increase in exposure), lower rates of tertiary educational attainment (-24.70 percentage points), and lower participation in non-formal training (-8.55 percentage points).

Keywords: artificial intelligence, occupational exposure, workforce preparedness, adult learning, PIAAC

1. Introduction

Artificial intelligence (AI) has revived a long-running question about technological change and work: when machines become capable of activities previously carried out by people, which workers are positioned to adapt? Most of the debate begins with occupational exposure, meaning the extent to which AI capabilities overlap with the tasks an occupation requires. Exposure, however, is neither adoption nor an employment outcome. Task-based research shows that technologies alter bundles of tasks and can displace, complement and create human work at the same time, while realized consequences depend on organizational choices, complementary investment and institutions (Autor et al., 2003; Acemoglu & Restrepo, 2019).

A large part of the conversation slides from exposure to prediction. Exposure indices are widely reported as though they ranked jobs by probability of disappearance, a reading their authors generally reject. Frey and Osborne (2017) were quickly read as forecasting job losses, whereas task-level reanalysis produced substantially smaller figures once within-occupation heterogeneity was allowed (Arntz et al., 2016). Exposure measures describe a technological potential whose consequences are settled elsewhere, in workplaces, training systems and institutions.

This paper therefore asks a question that is prior to adoption. Before AI systems are widely deployed in the occupations a capability-based index identifies as exposed, what capabilities and learning resources do the workers already in those occupations possess? The answer is not obvious. If exposure concentrates in professional and analytically intensive work, exposed occupations should contain workers with stronger measured proficiency and more education. If it concentrates in routine information-processing and codifiable work, the pattern may reverse. Because different indices encode different technological frontiers, the alignment between exposure and preparedness has to be established for each measure, not assumed.

2. Theoretical background and research methodology

2.1. Task-based accounts and the social shaping of outcomes

The dominant vocabulary for technological change at work is task-based. Instead of treating technology as raising or lowering demand for skill in general, this tradition disaggregates occupations into bundles of tasks and asks which tasks machines can perform (Autor et al., 2003; Acemoglu & Autor, 2011). Its central insight is that automation displaces and creates tasks simultaneously, so aggregate employment effects depend on the balance between the two (Acemoglu & Restrepo, 2019). Two limitations matter here. Tasks are implemented not by machines but by organizations: Bresnahan et al. (2002) showed that returns to information technology depended on complementary organizational redesign and human capital, and that firms investing without the complements did not obtain the gains. And technical feasibility is a poor predictor of whether a task will be automated, because feasibility says nothing about relative cost, regulation, worker resistance or managerial preference.

A second tradition insists that technological outcomes are shaped, not given. Braverman (1974) treated technology as an instrument of managerial control. Barley (1986) demonstrated the point directly: two radiology departments introduced identical CT scanners and arrived at different occupational structures, because the technology functioned as an occasion for structuring rather than as its cause. Orlikowski (2000) generalized the argument through a practice lens in which structures emerge from recurrent use, and Howcroft and Taylor (2023) extend the critique to automation narratives. Holm and Lorenz (2022) show that realized AI use altered skill requirements differently depending on occupation and mode of use, which a deterministic reading of exposure cannot accommodate. If outcomes are indeterminate, studying the conditions that enter the shaping process is more informative than predicting outcomes from capability.

2.2. Preparedness as occupational absorptive capacity

To specify preparedness the study borrows, with modification, the concept of absorptive capacity. Cohen and Levinthal (1990) argued that an organization's ability to recognize, assimilate and apply new external knowledge depends on its prior related knowledge, and that the capacity is cumulative. The concept was developed for firms, and occupations are not organizations; what transfers is the structural intuition. The capacity of a group of workers to absorb a technological change depends on what they already know, on the credentials that give them access to redesigned roles, and on whether they are currently learning at all.

This yields three dimensions. Adaptive problem solving captures the general capacity to make progress on unfamiliar, ill-structured problems. Tertiary attainment captures accumulated and credentialled human capital, which governs mobility into reconfigured roles. Participation in non-formal education or training captures whether learning is currently active, the dimension most amenable to policy and employer action. Adult learning is also where inequality is most firmly established: participation is strongly stratified by prior education, occupational position and employer support, producing a documented Matthew effect (Desjardins & Rubenson, 2013; Boeren, 2016). If exposure concentrates among workers with weaker initial education, the same stratification implies they will also be least likely to receive the training that adaptation would require.

2.3. Competing exposure frontiers and the research gap

Exposure measures differ in what they measure, and the differences are consequential. Felten et al. (2021) map progress in AI applications onto occupational abilities; Eloundou et al. (2024) estimate the share of occupational tasks whose completion time large language models could substantially reduce. Both identify considerable exposure in highly educated, language-intensive professional work. The refined ILO index likewise finds high exposure in clerical occupations while emphasizing transformation over replacement (Gmyrek et al., 2025).

The OECD measure introduced in 2026 asks a different question. It maps nine cognitive, social and physical AI capability domains onto occupational requirements and constructs an AI Capability Gap; smaller gaps mean current systems sit closer to the capability profile an occupation requires, and therefore indicate greater potential exposure (OECD, 2026). The index used here reverses that gap orientation solely so higher values consistently mean greater exposure. Under this frontier, exposure is especially associated with routine information-processing, administrative and codifiable work, whereas occupations requiring contextual judgement, interpersonal understanding and complex responsibility retain larger gaps. A positive relationship between education and exposure observed under other indices therefore cannot be assumed under this measure.

Cross-national work has already linked worker characteristics to occupational AI exposure (Georgieff & Hyee, 2021; Lewandowski et al., 2026), so linking PIAAC to exposure is not novel in itself. The narrower gap addressed here is the joint use of Cycle-2 direct proficiency, educational attainment and adult learning as dimensions of incumbent preparedness for the new OECD capability-gap frontier, combined with a robustness architecture that tests occupational resolution, country weighting, country influence and occupation-level dependence.

2.4. Hypotheses

Three positive hypotheses were prespecified before the main results were examined, on the reasoning that analytically demanding occupations tend to be both more technologically intensive and more highly skilled. H1 predicted that greater occupational AI exposure would be associated with higher adaptive problem-solving proficiency; H2 predicted a higher probability of tertiary attainment; H3 predicted a higher probability of recent non-formal education or training. They are retained in their original directions despite the observed results. Reversing them after the fact would misdescribe what was actually predicted.

2.5. Data and sample

The study is a cross-national secondary analysis linking worker characteristics in the OECD 2023 Survey of Adult Skills (PIAAC Cycle 2) to the subsequent OECD 2026 occupational AI exposure measure. PIAAC data were collected primarily in 2022–2023 and provide population-based measures of proficiency, education, learning, employment and demographics (OECD, 2025). The primary sample was restricted to employed adults aged 25–65 with a valid two-digit ISCO-08 exposure match, giving 52,032 respondents from Belgium (Flemish Region), Chile, Czechia, Spain, France, Hungary, Israel, Italy, Japan, Korea, Latvia, Poland, Portugal, Singapore, the Slovak Republic and the United States. Country inclusion and merge rules were frozen before substantive modelling. A prespecified validation sample supported four-digit exposure estimation in seven countries (N = 22,217), of whom 22,145 had both valid two- and four-digit exposure.

2.6. Measures

The primary exposure variable is a frozen 0–1 occupation-level transformation of the OECD AI Capability Gap, audited so that higher values unambiguously indicate greater estimated exposure; all effects are reported for a +0.10 increase. The measure denotes neither observed AI adoption, automation probability nor job-loss risk. At the top of the distribution sit general and keyboard clerks (0.983), other clerical support workers (0.969) and numerical and material recording clerks (0.954), which makes the construct's occupational meaning concrete. Adaptive problem solving was the primary capability outcome and used all ten PIAAC plausible values; literacy and numeracy were secondary domains. Tertiary attainment equalled 1 for ISCED 5–8. Recent adult learning was participation in non-formal education or training during the previous 12 months. Covariates were country fixed effects, 10-year age group and gender, with education added only where prespecified.

The linkage is where measurement error enters. Each respondent was assigned the exposure value of their reported occupational unit, so all workers sharing a code receive an identical score regardless of task profile, employer or technology in use. This carries a standard cost: within-occupation heterogeneity goes to the residual, and standard errors treating respondents as independent will be too small. Both problems were addressed prospectively, through four-digit validation and occupation-cluster inference.

2.7. Estimation and robustness plan

All models used the PIAAC final weight and the effective replicate weights, preserving each country's native JK2 or Fay BRR design. The primary pooled estimand gave each country equal total weight using fixed senate-style factors frozen before modelling; population-weighted results were treated as a distinct robustness estimand and not as a correction, because the two answer different questions. Because countries use different replication factors, pooled inference used 1,164 normalized source-stacked perturbations. Models were estimated separately for all ten plausible values and combined following PIAAC analysis guidance (OECD, 2025).

The model hierarchy was frozen prospectively. A0 regressed adaptive problem solving on continuous exposure and country fixed effects; A1, the H1 model, added age and gender; A2 added education and A3 added non-formal training, with coefficient changes across A1–A3 defined as sequential compositional adjustments and not mediation. H2 and H3 used survey-weighted logistic models with Holm adjustment, reporting average marginal probability differences per +0.10 exposure. Six robustness tests assessed occupational resolution, population weighting, country-specific slopes, leave-one-country-out influence, and occupation-level dependence using CR2 cluster-robust inference across 40 two-digit occupational clusters with a restricted wild-cluster bootstrap. Analyses were implemented in Python with frozen inputs and reproducible seeds; missingness was handled by model-specific eligibility without imputation.

3. Results and discussion

3.1. Descriptive composition

Under equal-country weighting, mean occupational exposure was 0.852, mean adaptive problem-solving proficiency was 246.70, 44.9% of workers were tertiary educated and 42.3% had participated in non-formal training. Unadjusted exposure quartiles already showed a strong compositional gradient (Table 1). Two features qualify it. The decline is not strictly monotonic: tertiary attainment was 23.9% in Q3 and 25.3% in Q4, so the gradient is driven by the contrast between the two lower and the two upper quartiles. And gender composition varies sharply and non-monotonically, from 57.0% female in Q1 to 16.9% in Q3 and 61.7% in Q4. Q3 is dominated by male-typed manual, craft and operator occupations, whereas Q4 combines clerical and service work in which women are heavily over-represented.

Table 1. Sample characteristics by occupational AI-exposure quartile

Source: Author's calculations from PIAAC Cycle 2 (OECD, 2025) and the OECD AI exposure measure (OECD, 2026). APS = adaptive problem solving. N is unweighted; all other figures use equal-country weighting.

3.2. Confirmatory and sequential results

All three prespecified positive hypotheses were unsupported (Table 2). A +0.10 increase in exposure was associated with -11.886 proficiency points (SE = 0.372, 95% CI [-12.615, -11.158], p < 0.001). The average marginal difference in tertiary attainment was -24.695 percentage points (SE = 0.320, Holm-adjusted p < 0.001), and in non-formal training participation -8.545 percentage points (SE = 0.369, Holm-adjusted p < 0.001). The interquartile range of exposure spans roughly 0.18 index units, so the coefficient implies a difference of about 21 proficiency points between the first and third quartiles, comparable in magnitude to several estimated country fixed effects.

Sequential models indicate substantial compositional overlap with education and learning. Adding education reduced the absolute coefficient from -11.886 to -3.637, an attenuation of 69.4%; adding non-formal training reduced it further to -2.535, or 78.7%. Weighted R-squared rose from 0.172 in A1 to 0.271 once education entered and 0.285 after training. These are not mediation estimates and do not establish that education causally explains the gradient. What they establish is that the gradient and the educational stratification of the workforce are largely the same phenomenon measured two ways, which is itself the substantive point.

Table 2. Main and sequential model results

Source: Author's calculations. AME = average marginal effect. Coefficient differences across A1–A3 are sequential compositional adjustments, not mediation estimates. Holm adjustment applied to H2 and H3.

3.3. Robustness

The inverse association survived every core test (Table 3). Four-digit exposure produced -8.119, and on the identical 22,145-person sample two-digit and four-digit estimates were -11.687 and -8.129; the roughly 30% reduction at finer resolution suggests two-digit aggregation overstates the gradient somewhat while leaving direction and significance unaffected. Population-weighted estimates were -13.659 for proficiency, with tertiary and training differences also negative. Leave-one-country-out estimates ranged from -12.436 to -11.308 with no sign reversal, the largest single-country influence being 0.578 points. CR2 inference across 40 occupational clusters widened the standard error to 3.157 but retained a negative interval of [-18.075, -5.698] with p = 0.006, and the wild-cluster bootstrap gave p = 0.001. That interval is roughly eight times wider than the survey-design interval, a reminder that the effective number of independent occupational units is small.

Country-specific slopes were negative in all 16 countries, with every confidence interval below zero, though magnitudes differed nearly fivefold, from -21.097 in Chile to -4.413 in Korea. Secondary capability outcomes were also inverse: -16.180 literacy points and -17.214 numeracy points, both Holm-adjusted p < 0.001. Exploratory moderation was not established (education interaction p = 0.054; training interaction p = 0.692).

Table 3. Core robustness of the inverse preparedness gradient

Source: Author's calculations. The population-weighted model is a robustness estimand, not a replacement for the primary equal-country estimate. The restricted wild-cluster bootstrap gives p = 0.001.

3.4. Discussion

The result is less paradoxical once the exposure construct is separated from generic skill intensity. Several influential indices emphasize language-intensive, professional and highly educated work (Felten et al., 2021; Eloundou et al., 2024). The OECD 2026 measure asks how close current AI capabilities are, across nine domains, to the capability requirements of occupations, and its smaller gaps are often found in routine information-processing work (OECD, 2026). The negative gradients therefore cannot be dismissed as error merely because they differ from results produced by other indices. They are measure-specific evidence about a different technological frontier, and the divergence between indices is a finding about the construct, not a problem to be reconciled away.

This is also where the social shaping tradition bears on the interpretation. The analysis measures potential occupational exposure and not realized use; work design, managerial strategy, worker voice, training and regulation all remain outside the exposure score (Howcroft & Taylor, 2023; Holm & Lorenz, 2022). The inverse gradient identifies unequal starting conditions that organizations and institutions may amplify, mitigate or transform. It does not specify the eventual direction of work change, and it should not be read as predicting that lower-preparedness workers will necessarily lose. Brynjolfsson et al. (2025) show that generative AI can disproportionately raise productivity among less experienced workers in a specific organizational setting. Baseline preparedness may signal vulnerability where learning and worker voice are weak, but it may equally indicate room for augmentation where technology is introduced with support.

Continuing training is particularly relevant because adaptation depends not only on baseline knowledge but on access to capability renewal, and that access is itself unequally distributed (Desjardins & Rubenson, 2013; Boeren, 2016). The occupational groups closest to this frontier are, on average, those least engaged in recent learning: if the Matthew effect operates as documented, the mechanism that would ordinarily close a preparedness gap is weakest exactly where the gap is widest. The gender pattern deserves the same care. Because the highest-exposure quartile is disproportionately female and clerical, the inverse gradient carries a gender dimension that the capability-gap framing makes newly visible (Lane, 2024; Charles & Grusky, 2004). Gender is adjusted for in every confirmatory model and the gradient survives, but adjustment is not analysis.

The common negative sign across 16 countries should not be translated into a claim of homogeneous effects. A purely institutional account would predict that the association appears in some national contexts and not others, since training regimes and occupational structures differ substantially. That is not what the data show. The direction is common and the magnitude varies, which is more consistent with the gradient being carried by the occupational structure the exposure measure encodes, with institutions modulating its intensity but not determining its sign.

4. Conclusion

Under the OECD 2026 measure, and across 16 countries, the workers already occupying occupations near the AI capability frontier do not possess the capabilities and learning resources that adaptation would require. Proficiency, tertiary attainment and recent training all decline as estimated exposure rises, and the proficiency gradient is robust to occupational resolution, weighting, country omission and occupation-cluster inference. The result is associational and measure-specific, and it says nothing about whether these workers will be displaced, augmented or left untouched.

For employers, the implication is that AI readiness assessments built on exposure scores alone will systematically misidentify where support is needed. An organization that ranks its occupations by exposure and assumes those groups are its most capable will invest in the wrong places. Exposure audits should be paired with direct assessment of incumbent capability and current training participation at the level of the occupational group. Consultation and work design matter more, not less, where preparedness is low, since these are the groups least able to negotiate the terms of redesign from a position of scarce expertise.

For policy, the results bear on the targeting of adult learning provision. If publicly and employer-funded training continues to follow existing educational stratification, it will flow away from the occupations this measure identifies as most exposed. Compensatory instruments — individual learning entitlements weighted by prior attainment, sectoral training funds in clerical and administrative work, and collectively negotiated paid release — address the documented pattern in a way that general exhortation to upskill does not. The cross-national spread additionally implies that appropriate policy intensity is not uniform: a gradient of -21 points and one of -4 points are not the same problem.

Several limitations bound these conclusions. Occupational self-selection, educational sorting and unobserved job characteristics can generate the observed associations, and no part of the design supports a causal reading. Exposure is assigned at occupational level, not at workplace or individual level, and the roughly 30% attenuation at four-digit resolution indicates that aggregation is not innocuous. The temporal structure maps 2022–2023 worker characteristics to a subsequent 2026 frontier, which protects against claims that AI caused lower preparedness but makes the study a retrospective baseline mapping. PIAAC measures general information-processing skills, not AI literacy, and training participation says nothing about quality, duration or employer sponsorship. The gender pattern was not part of the frozen analysis plan. Finally, the 16-country sample is skewed toward higher-income economies. Before asking who will adapt, it is worth knowing who is currently in a position to.

Reference:

Acemoglu, D., & Autor, D. (2011). Skills, tasks and technologies: Implications for employment and earnings. In O. Ashenfelter & D. Card (Eds.), Handbook of Labor Economics (Vol. 4B, pp. 1043 - 1171). Elsevier. https://doi.org/10.1016/S0169-7218(11)02410-5

Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3 - 30. https://doi.org/10.1257/jep.33.2.3

Arntz, M., Gregory, T., & Zierahn, U. (2016). The risk of automation for jobs in OECD countries: A comparative analysis (OECD Social, Employment and Migration Working Papers No. 189). OECD Publishing. https://doi.org/10.1787/5jlz9h56dvq7-en

Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. Quarterly Journal of Economics, 118(4), 1279 - 1333. https://doi.org/10.1162/003355303322552801

Barley, S. R. (1986). Technology as an occasion for structuring: Evidence from observations of CT scanners and the social order of radiology departments. Administrative Science Quarterly, 31(1), 78 - 108. https://doi.org/10.2307/2392767

Boeren, E. (2016). Lifelong learning participation in a changing policy context: An interdisciplinary theory. Palgrave Macmillan.

Braverman, H. (1974). Labor and monopoly capital: The degradation of work in the twentieth century. Monthly Review Press.

Bresnahan, T. F., Brynjolfsson, E., & Hitt, L. M. (2002). Information technology, workplace organization, and the demand for skilled labor: Firm-level evidence. Quarterly Journal of Economics, 117(1), 339 - 376. https://doi.org/10.1162/003355302753399526

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889 - 942. https://doi.org/10.1093/qje/qjae044

Charles, M., & Grusky, D. B. (2004). Occupational ghettos: The worldwide segregation of women and men. Stanford University Press.

Cohen, W. M., & Levinthal, D. A. (1990). Absorptive capacity: A new perspective on learning and innovation. Administrative Science Quarterly, 35(1), 128 - 152. https://doi.org/10.2307/2393553

Desjardins, R., & Rubenson, K. (2013). Participation patterns in adult education: The role of institutions and public policy frameworks in resolving coordination problems. European Journal of Education, 48(2), 262 - 280. https://doi.org/10.1111/ejed.12029

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306 - 1308. https://doi.org/10.1126/science.adj0998

Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, 42(12), 2195 - 2217. https://doi.org/10.1002/smj.3286

Frey, C. B., & Osborne, M. A. (2017). The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change, 114, 254 - 280. https://doi.org/10.1016/j.techfore.2016.08.019

Georgieff, A., & Hyee, R. (2021). Artificial intelligence and employment: New cross-country evidence (OECD Social, Employment and Migration Working Papers No. 265). OECD Publishing. https://doi.org/10.1787/c2c1d276-en

Gmyrek, P., Berg, J., Kaminski, K., et al. (2025). Generative AI and jobs: A refined global index of occupational exposure (ILO Working Paper No. 140). International Labour Organization. https://doi.org/10.54394/HETP0387

Holm, J. R., & Lorenz, E. (2022). The impact of artificial intelligence on skills at work in Denmark. New Technology, Work and Employment, 37(1), 79–101. https://doi.org/10.1111/ntwe.12215

Howcroft, D., & Taylor, P. (2023). Automation and the future of work: A social shaping of technology approach. New Technology, Work and Employment, 38(2), 351 - 370. https://doi.org/10.1111/ntwe.12240

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. https://doi.org/10.1787/14dc6f89-en

Lewandowski, P., Madon, K., & Park, A. F. (2026). Workers' exposure to artificial intelligence across development stages (ADB Economics Working Paper Series No. 836). Asian Development Bank. https://doi.org/10.22617/WPS260018-2

OECD. (2025). Survey of Adult Skills 2023 data analysis manual. OECD Publishing. https://doi.org/10.1787/25a87a9d-en

OECD. (2026). The OECD AI exposure measure: Mapping the OECD AI Capability Indicators to occupations (OECD Artificial Intelligence Papers No. 59). OECD Publishing. https://doi.org/10.1787/f3da0f0a-en

Orlikowski, W. J. (2000). Using technology and constituting structures: A practice lens for studying technology in organizations. Organization Science, 11(4), 404 - 428. https://doi.org/10.1287/orsc.11.4.404.14600

Mức độ tiếp xúc nghề nghiệp với trí tuệ nhân tạo và sự sẵn sàng của người lao động: Kỹ năng, trình độ học vấn và học tập của người trưởng thành tại 16 quốc gia

Lưu Ngọc Kỷ

Giảng viên, Học viện Hàng không Việt Nam, TP.Hồ Chí Minh, Vietnam

Tóm tắt:

Các tranh luận về trí tuệ nhân tạo (AI) trong môi trường làm việc phần lớn tập trung vào việc xác định những nghề nghiệp chịu tác động nhiều nhất, trong khi ít chú ý hơn đến mức độ sẵn sàng thích ứng của những người lao động hiện đang làm việc trong các nghề đó. Nghiên cứu này sử dụng dữ liệu khảo sát trên 52.032 người lao động trong độ tuổi 25 - 65 tại 16 quốc gia từ Khảo sát Kỹ năng Người trưởng thành năm 2023 của OECD với thước đo mức độ tiếp xúc nghề nghiệp với AI của OECD năm 2026. Trái với giả thuyết đặt ra ban đầu, mức độ tiếp xúc cao hơn gắn với năng lực giải quyết vấn đề thích ứng thấp hơn (-11,89 điểm trên mỗi mức tăng 0,10), tỷ lệ có trình độ đại học trở lên thấp hơn (-24,70 điểm phần trăm) và tỷ lệ tham gia đào tạo phi chính quy thấp hơn (-8,55 điểm phần trăm). Xu hướng nghịch này ổn định ở cả 16 quốc gia.

Từ khóa: trí tuệ nhân tạo, tiếp xúc nghề nghiệp, mức độ sẵn sàng của lực lượng lao động, học tập của người lớn, PIAAC

[Tạp chí Công Thương - Các kết quả nghiên cứu khoa học và ứng dụng công nghệ, Số 24 năm 2026]