Augmentation AI is creating new, higher-paid work and drawing students into exposed fields, while automation AI lowers wages and worsens outcomes for low-skilled workers; employers are increasingly hiring AI, data and prediction skills, and returnees from cross-border work suffer a short-lived employment penalty and higher unemployment claims.

Main Finding

This PhD dissertation (David Marguerit, 2026) provides a multi‑method empirical assessment of how recent AI development reshapes work, skill demand, education choices, and cross‑border labor flows. Using novel measures of occupational AI exposure (distinguishing automation vs. augmentation), linked administrative and online data, instrumental‑variable identification, and causal machine‑learning, the thesis shows that AI is simultaneously creating "new work" and re‑shaping employment and wages heterogeneously across occupations and skill groups. AI exposure shifts demand toward cognitive/technical and complementary skill bundles, alters college major choice and institutional responses, and interacts with labor mobility in ways that matter for post‑return labor‑market outcomes.

Key Points

• Distinguishing automation vs. augmentation matters: the dissertation constructs separate "Automation AI" and "Augmentation AI" exposure indices (built from Stack Overflow activity and validated with other sources) and demonstrates they have distinct empirical relationships with tasks, job titles, and outcomes.
• New tasks and new job content emerge as AI develops: occupations with higher augmentation exposure show more creation of new task content and alternate job titles over 2015–2022.
• Heterogeneous labor effects:
  • AI exposure is associated with reallocation of employment across occupations and within‑occupation task changes. Effects on employment and wages are heterogeneous by occupation and skill group.
  • Instrumental variable (CS publications) 2SLS estimates are used to address endogeneity of AI adoption, yielding stronger causal claims about how AI exposure affects new work, employment (log employment), and mean wages.
• Skills demand shifts:
  • Analysis of European online job vacancies (2018–2023) using LLM‑based skill extraction shows AI exposure increases demand for certain cognitive/technical skills and bundles that are complementary to AI, while reducing demand for skill bundles that are closer substitutes to automation.
  • Theoretical model links AI exposure to production hierarchy and predicts within‑occupation re‑weighting of skill tasks; empirical results broadly support these predictions.
• Human capital and education responses:
  • Occupational AI exposure influences college major selection. Students update intentions and choices in response to AI exposure; higher AI automation risk in an occupation is associated with measurable shifts in field‑of‑study outcomes at the bachelor level.
  • Educational institutions also respond (openings/closures of programs, curriculum changes) consistent with changing demand signals.
• Labor mobility and AI interactions:
  • Using administrative data for Luxembourg, causal machine‑learning methods (prediction and causal ML) identify key predictors of initiating and returning from cross‑border work (CBW) and quantify consequences for post‑return labor‑market states, job quality, and unemployment benefit receipt.
  • Cross‑border experience is an important margin of adjustment with heterogeneous returns depending on worker characteristics and history.

Data & Methods

• AI exposure construction:
  • Primary innovation: two occupational AI exposure indices (Automation AI and Augmentation AI) derived from technical activity and tagging on Stack Overflow (2010–2022), validated against firm‑level AI adoption and other external measures.
  • Instrument: exposure to computer‑science (CS) scientific publications used to construct IVs (regional/occupational shocks in CS R&D).
• Labor market outcomes:
  • Employment counts, new task measures (new titles/alternate titles), and wages analyzed at occupation × time granularity; control variables include industry, region, and task content (O*NET) measures.
• Skill demand analysis:
  • Large panel of online job vacancies across European countries (Lightcast/other vacancy sources), with LLM‑based extraction and classification of skills into granular skill groups and broader classes; weighting schemes to map vacancies to occupations/industries/locations.
  • Instrumental‑variable strategy: local exposure to CS publications as an instrument for occupational AI exposure.
• Human capital analysis:
  • Administrative/education datasets on bachelor graduates, intended fields, SAT scores, and program openings/closures; occupational AI exposure mapped into field‑of‑study exposure (state/field level).
  • Identification via panel variation and IVs where appropriate; analysis of both intentions and realized field choice.
• Cross‑border work:
  • Luxembourg administrative labor‑market data covering worker demographics, employment history, cross‑border links.
  • Machine‑learning toolkit: XGBoost and causal ML methods (causal forests / doubly robust learners) for prediction of CBW start/return and for estimating causal effects of returning on subsequent outcomes; extensive validation and partial‑dependence/variable‑importance diagnostics.
• Robustness and validation:
  • Extensive appendices report robustness checks: balanced panels, different weighting schemes, outlier removal, alternative tag descriptions, full transition matrices, OLS vs 2SLS, first‑stage diagnostics, and validation of LLM skill extraction against human benchmarks.

Implications for AI Economics

• Measurement contribution: separating automation vs. augmentation exposures (and providing validated indices) improves precision in empirical studies of AI’s labor effects and should be adopted in future research and policy monitoring.
• Policy on skills and training:
  • Evidence that AI shifts within‑occupation task mixes and increases demand for complementary cognitive/technical skills argues for targeted upskilling/reskilling programs focused on augmentation complementarities rather than one‑size‑fits‑all approaches.
  • Early signals of student responsiveness to AI exposure imply education policy (career guidance, curriculum redesign) can materially affect the supply of human capital aligned with evolving demand.
• Labor‑market adjustment and mobility:
  • Cross‑border work is an important margin of adjustment; policies facilitating mobility, recognition of cross‑border experience, and portability of social insurance can influence workers’ ability to adapt to AI‑driven change.
• Distributional concerns:
  • Heterogeneous employment and wage effects across occupations/skill groups highlight potential inequality risks. Safety nets, active labor‑market policies, and progressive upskilling investments are needed to mitigate adverse distributional impacts.
• Research and evaluation:
  • The dissertation’s combination of new measurement, causal identification, and ML methods offers a template for continued evaluation of AI’s economic effects. Policymakers and researchers should integrate high‑frequency indicators (e.g., technical forum activity, vacancy text analytics) into monitoring frameworks to detect and respond to rapid AI‑driven labor shifts.

If you want, I can (a) extract the specific main empirical coefficients and quantitative magnitudes from the thesis tables (e.g., Tables 1.2–1.5, 2.5.1, 3.4.3, 4.5.*) and summarize them in a short numeric table, or (b) produce a two‑page policy brief tailored for labor‑market or higher‑education policymakers. Which would you prefer?