Rapid advances in technology and events such as COVID-19 have significantly transformed the modern workplace, potentially altering the skills demanded in jobs. This study examines the evolving demand and posted-wage premia for Information and Communication Technology (ICT), interpersonal, and Artificial Intelligence (AI) skills in the UK labour market. Using a comprehensive dataset of online job advertisements (2016 to 2022), skills are extracted and categorised via GPT-4 zero-shot learning. Cross-sectional log-wage regressions, incorporating occupation and regional fixed effects with three-way clustered standard errors, reveal divergent trends in skill compensation. While interpersonal skills are ubiquitously demanded (approximately 90% of listings), they yield no significant posted-wage premium. In contrast, ICT skills, demanded in approximately 55% of postings, carry a posted-wage premium of approximately 7%. AI skills, mentioned in approximately 3% of postings, carry a posted-wage premium of approximately 9% within the ICT-mentioning subsample.

A widespread view holds that the adoption of large language models (LLMs) follows capability at two levels: models are used for the tasks they perform well (machine capability), and the regions whose own capabilities match those tasks (regional capabilities) use them most and take up each improvement first. We test this view on four waves of US state-level data on Anthropic’s Claude, sampled between August 2025 and May 2026, spanning three releases that improved coding. Skill-based measures of regional capabilities built from job postings are the main regressors, alongside the occupational composition of state employment. Regional capabilities are strongly associated with how intensively a state uses the model. Much of this is shared with income and GDP, but skill relatedness survives every control, and in the 2026 data it also appears in Microsoft’s estimates of AI users across assistants. Regional capabilities are not associated with what the model is used for: use follows local occupational composition only weakly, and no more strongly where regional capabilities are deeper. Nor are they associated with how augmentatively states use the model, or with which states gained share of use after the releases; nationally, coding lost more than a third of its share while education and business tasks gained. Because more than nine in ten state–task counts are unpublished, the extensive margin is not identified, and an imputation that fits the published counts overstates the unpublished ones.

We construct the measurement inputs for the empirical analysis of the digital and green transitions in European labour markets: technology taxonomies, exposure measures, and a measure of substitution and complementarity. A digital taxonomy of 9 domains and 184 technologies is mined from Wikipedia and 1.9 million patents; the green taxonomy adopts the OECD ENV-TECH classification. Exposure of 427 occupations and 272 industries is measured identically in both domains as citation-weighted semantic similarity between task and patent text, validated against ten published exposure indices and by inspection of the full rankings. Text-matched digital tagging requires corrective cleaning of 1.3 percent of cells, examiner-classified green tagging 0.06 percent, isolating the source of measurement error. Substitution and complementarity are elicited separately from a language-model panel for 11,564 technology-occupation-industry cells with pairs observed in European employment data; complementarity dominates, and the ratings are read as fractions of the reached task content and combined with exposure into substituted and augmented content, two levels whose mix varies across technologies and occupations but not across industries. All rejected alternatives are documented.

Can a language model reproduce what domain experts think? We test this against a completed Delphi survey in which human experts rated how specific technologies will change work tasks, regenerating every response with a persona-conditioned model call. At the aggregate level, yes: after screening eight models, the selected model reproduces the human experts’ rating distribution with a similarity of 0.945. The agreement is between distributions, not individuals. Similarity declines as the distributions are compared within finer groups of cases, from 0.93 for a whole construct to 0.78 within a single technology, industry and task; it is insensitive to the respondent’s profile; and it rises with the homogeneity of the human ratings: where the experts’ ratings agree, the silicon experts match them; where they divide most, on substitutability, the model rates replacement lower. A silicon panel stands in for the aggregate judgement of experts, not for the expert.

Standard Occupational Classifiers (SOC) are systems used to categorize and classify different types of jobs and occupations based on their similarities in terms of job duties, skills, and qualifications. Integrating these facets with Big Data from job advertisement offers the prospect to investigate labour demand that is specific to various occupations. This project investigates the use of recent developments in natural language processing to construct a classifier capable of assigning an occupation code to a given job advertisement. We develop various classifiers for both UK ONS SOC and US O*NET SOC, using different Language Models. We find that an ensemble model, which combines Google BERT and a Neural Network classifier while considering job title, description, and skills, achieved the highest prediction accuracy. Specifically, the ensemble model exhibited a classification accuracy of up to 61% for the lower (or fourth) tier of SOC, and 72% for the third tier of SOC. This model could provide up to date, accurate information on the evolution of the labour market using job advertisements.

Technological change has an ambiguous impact on labour market by creating demand for some skills and reducing demand of some others. Hence the relationship of demand for skills and the need for training is an empirical question. In this paper, I investigate association between types of technological change and decisions. I categorise several measures of automation on the basis of tasks (done at individual or occupation level) and technology (Software, Robot, AI) and compare their relationship with human capital investment. I find that the correlation between automation and training varies depending on the automation measure used, showing a decline with both individual- and occupation-level automation measures. However when relying on technology based measures of automation, workers exposed to older technologies (Robot, Software) receive less training with automation, while workers exposed to newer technologies (AI) receive more training with automation. The findings are consistent across workers of different age groups and skill levels.

Where does a new general-purpose technology take root, and why do some regional labor markets absorb it faster than others? We track firm AI adoption at biweekly frequency across US states and large metros in the Census Business Trends and Outlook Survey, and relate it to each region’s pre-existing skill structure, measured from 433.6 million job postings. Three findings emerge. First, the geography of adoption is largely pre-determined: skill structure alone predicts a fully held-out region’s adoption with a correlation of 0.63, and the 2019 skill base predicts 2023–25 adoption nearly as well as the 2024 base. Second, the absorption margin is the connective tissue of the local skill ecosystem (its relatedness density), not its sophistication: one standard deviation of density predicts about one percentage point of additional adoption growth, while skill complexity predicts none. Third, each frontier model release widens the adoption gap between exposed and unexposed regions by roughly 0.1 percentage points, a treadmill that compounds; this dynamic reading is the hardest to separate from a smooth differential trend. Because laggards converge only to the steady states their fundamentals imply, a readiness typology of regions, and the margin each should target, follows.

We link patent-based measures of green and digital technological exposure to the skill content of European online job vacancies, at the regional level, across occupations, and within occupations. Digital exposure is strongly associated with greater demand for digital skills and with higher skill demand overall; green exposure is associated with the demand for green skills at the occupational level; and joint exposure to both technologies displays a distinct pattern that cannot be inferred from either transition in isolation.

This paper explores how evolving regional skill demand relates to labour productivity. It draws on over 170 million Lightcast online job advertisements, at the NUTS-3 level across the UK and ten EU countries, between 2014 and 2019. Categorising UK skills with the Lightcast taxonomy and EU skills with the ESCO framework, we measure regional skill demand by network-based coherence, fitness, and entropy of skills sets. Combined with ARDECO productivity data and socioeconomic controls, we apply two-way fixed effects and long-difference regressions. For EU regions, more diverse and less complex skill demand is positively associated with productivity, while UK regions suggest the opposite. This study presents the first systematic evidence on the role of evolving skill demand structures in shaping regional productivity dynamics across Europe and the UK.

This paper examines how the twin transition relates to labour productivity by tracing the evolving regional demand for green and digital skills. We draw on more than 130 million Lightcast online job advertisements across ten European countries at the NUTS-3 level between 2014 and 2019, using the ESCO framework to classify green and digital skills. Regional skill-demand shares are combined with ARDECO productivity data and socioeconomic controls and analysed using two-way fixed-effects and long-difference regressions, complemented by a shift-share instrumental-variable approach. We show that greening and digitalisation are predominantly within-industry dynamics. While demand for digital skills is positively associated with regional labour productivity and allows for causal inference, demand for green skills is only negatively correlated with it.

Skill taxonomies extracted from online job advertisements are often mutually incomparable; we map two Lightcast taxonomies into the public ESCO standard. Where vocabularies overlap (the 1,747 Lightcast EU skills), exact label matching links 73% of skills and embedding similarity with structured expert review completes the mapping. Where vocabularies are disjoint (the 35,950-skill Lightcast UK taxonomy), label-based methods fail, and we instead match categories by their composition: each Lightcast subcategory and each ESCO category is represented by the centroid of its constituent skill embeddings. Centroid matching places 92.5% of subcategories above the 0.6 similarity threshold, against 39.2% for category-name and 9.9% for description matching, and the approach applies to any crosswalk between proprietary and official classifications. We validate the mappings by their consequences for measurement: the regional portfolio measures built on them are robust to the matching decisions, preserving regional rankings throughout. Benchmarked against measures computed directly on the raw Lightcast UK taxonomy, the centroid mapping also preserves the benchmark’s regional rankings (Spearman correlation of 0.91) where an alternative crosswalk based on individual preferred labels reorders them (0.56), although the two agree in levels (Pearson correlations of 0.98 and 0.96); a crosswalk should therefore not be judged by level agreement alone.

We estimate the association between digital and green technology exposure and employment across European regions, and ask which technologies carry it, when, and for whom. Patent-based exposure measures, built symmetrically for the two domains, are combined with 2009 industry structures into regional shift-share exposures and related to employment changes in 212 NUTS-2 regions of 18 countries over 2009–2019, 2009–2023 and 2020–2023, with the 2021 break in the Labour Force Survey corrected. Inference treats the ten industry shocks as the unit of randomness. One standard deviation of digital exposure is associated with a 0.97 percentage point larger change in the employment-to-population ratio over 2009–2019 (exposure-robust standard error 0.23, significant at 1 percent); over 2009–2023 the coefficient is 1.31 and over 2020–2023 0.47, neither significant. Green exposure carries no significant association on any window. Eight of the nine digital technology groups carry positive and significant employment-rate associations over 2009–2019, and only robotics and manufacturing carries one over 2020–2023, at 10 percent; in the green domain, green ICT and smart grids are positive and greenhouse-gas capture negative over 2009–2019. The digital association is larger for the low-educated than for the medium-educated, by 1.54 percentage points over 2009–2019, significant at 5 percent. Placebo tests select the employment rate over employment growth as the outcome of record.