Faster access to citizenship boosts immigrant integration—raising women’s employment by nearly nine percentage points and earnings by over 20%—while a small Airbnb profile photo change unexpectedly widened Black–White demand gaps; separately, an LLM scan of >4m firm websites finds AI-active European firms jumped from about 1% in 2016 to nearly 12% in 2024, concentrated in larger, younger, skill‑intensive urban firms.

Main Finding

Julio Garbers’s dissertation comprises four applied labor-economics essays. The AI-relevant core finding (Chapter 4) is that web-based, LLM-driven measures can map firm-level AI adoption at scale across Europe and reveal a rapid, uneven diffusion: the share of AI-active firms in Belgium, France, Germany and Luxembourg rose from ~1% in 2016 to ~12% in 2024 (with acceleration after 2022). Adoption concentrates in larger, younger, knowledge-intensive urban firms; foundational Data skills are a necessary base for adoption while specialized AI skills are strong complements. Complementary results from other chapters highlight how institutions and platform design (citizenship policy, Airbnb UI) materially shape economic integration and market outcomes — showing the importance of measurement, incentives, and interface design when studying AI’s economic effects.

Key Points

• Chapter 4 (Mapping AI)
  • New firm-level indicator of AI adoption using an LLM applied to >4 million firm websites (Belgium, France, Germany, Luxembourg; 2016–2024).
  • Detects not only adoption but role in AI ecosystem and technology types.
  • Share of AI-involved firms: ~1% (2016) → ~12% (2024), faster growth post-2022.
  • Structural shift from specialized AI core providers toward broader adopters and application developers.
  • Adoption concentrated in large, young, knowledge-intensive firms in urban innovation clusters.
  • Skills: Data skills are foundational; specialized AI skills are strong complements.
  • Website-based adoption measure diverges meaningfully from industry exposure-based indices (actual adoption ≠ potential exposure).
• Chapter 3 (Airbnb design and bias)
  • A design change reducing host profile picture size increased Black–White demand disparity by 3.3 percentage points (~30% of pre-intervention gap).
  • Mechanism: reduced visibility of positive facial cues (e.g., smiles) led guests to rely more on skin color; Black hosts responded by updating photos and amenities.
  • Methods combined scraped platform data, fine-tuned face-classification (ViT), OLS, DiD, event-study and robustness checks.
• Chapter 2 (Citizenship & integration)
  • Faster access to citizenship in Germany improves immigrant economic integration, especially for women: female labor force participation +8.9 pp; earnings +21.3%; no significant effects for men.
  • Identification via local randomization and fuzzy RDD on residency requirement cutoffs.
• Chapter 1 is a survey synthesizing quasi-experimental evidence that faster/less onerous access to citizenship increases take-up and improves diverse integration outcomes.

Data & Methods

• Chapter 4 (AI mapping)
  • Data: >4 million firm websites (2016–2024) from four European countries; matched to Orbis firm registry for firm size, age, sector, location.
  • Approach: use a Large Language Model (LLM) / prompt-based pipeline to classify whether firms (and how) mention/offer/produce AI; extract role in AI ecosystem and technology categories. Skill mentions are parsed and mapped to a skill taxonomy; prompts and definitions documented in appendices.
  • Validation/representativeness: confusion matrix, cross-checks vs Orbis, balance tests; comparisons to industry-level exposure indices.
  • Empirics: time-series and cross-sectional regression analyses of determinants of adoption (controls: firm size, age, sector, region), and skill–adoption complementarity tests (regressions linking skill shares to adoption probability).
• Chapter 3 (Airbnb)
  • Data: scraped Airbnb listings in NYC across the design change; host pictures processed with a fine-tuned Vision Transformer (ViT) to predict race and facial cues.
  • Empirical strategy: OLS, difference-in-differences, event-study, spatial heterogeneity, robustness (SDiD, placebo tests).
• Chapter 2 (Naturalization)
  • Data: administrative / survey-linked data on first-generation immigrants in Germany.
  • Identification: exploitation of policy variation in residency requirements via Local Randomization around cutoffs and Fuzzy Regression Discontinuity Designs; outcomes include naturalization take-up, employment, earnings, benefits, heterogeneity analysis by gender.

Implications for AI Economics

• Measurement innovation
  • LLM-based scraping of firm websites provides high-frequency, firm-level indicators of AI adoption that complement (and often diverge from) industry-exposure indices. This enables more precise, micro-founded analyses of diffusion, productivity effects, and labor-market impacts.
  • Caveats: web-based measures capture self-reported/advertised activity — silent adopters or internal use not advertised may be missed; LLM classification risks false positives/negatives and requires validation.
• Policy targeting and labor markets
  • Findings that Data skills are necessary and specialized AI skills are complementary point to two-tier policy needs: broaden foundational data literacy across the workforce and target specialized AI training where complementary returns are highest.
  • Concentration of adoption in urban, knowledge-intensive firms suggests spatially targeted workforce and infrastructure policies to avoid uneven impacts and regional inequality.
• Research design for causal inference
  • The dissertation demonstrates the value of combining novel measurement (web/LLM, image classification) with quasi-experimental designs (RDD, DiD) to uncover causal pathways — an approach well-suited to studying AI’s heterogeneous effects on employment, wages, and firm performance.
• Platforms, UI and fairness
  • Chapter 3 shows how seemingly small UI changes can amplify or attenuate discrimination in AI-mediated markets. For AI economists, this implies careful attention to presentation and feature salience when assessing algorithmic marketplaces or deploying automated ranking/recommendation systems.
  • Policy/regulatory implications include UI audits, fairness-by-design, and monitoring downstream behavioral impacts of interface algorithms and platform updates.
• Directions for future work
  • Link web-derived AI adoption indicators to firm-level productivity, employment composition, wages, and innovation outcomes to quantify causal impacts of adoption.
  • Extend coverage beyond four countries and triangulate with procurement, patent, job-posting, and administrative data to capture silent adoption and validate LLM-based signals.
  • Study spillovers: how adoption by local firms or platform design changes affect labor demand for data vs. AI-specialized skills, and whether training programs can close the complementarity gap.
  • Explore equity and distributional consequences as adoption diffuses from core providers to widespread users.

Limitations to keep in mind: website/LLM methods reflect self-disclosure and marketing language; country sample limited to four European economies; image-based race inference raises ethical and measurement concerns (addressed in paper via robustness and validation but requires cautious use).

If helpful, I can (a) extract the Chapter 4 methodology and prompts in greater technical detail, (b) outline an empirical plan to link the website-based AI measure to firm-level productivity, or (c) produce reproducible pseudocode for the LLM-based classification pipeline. Which would you like next?