Generative AI is reshaping city labor markets by changing the tasks people do, not simply cutting jobs: fine-grained vacancy and firm-adoption data from 2018–2025 show skill bundles shifting toward collaboration and coordination as GenAI diffuses across cities.

Main Finding

Generative AI diffusion in cities causally shifts within-job skill bundles rather than primarily destroying jobs: high-GenAI cities see a sustained rise in complementary cognitive/coordination skills and a fall in routine skills. Effects accumulate over several quarters and are partly mediated by increased human–AI collaboration and cross-functional coordination tasks.

Key Points

• Data & scope: Unbalanced city–occupation–week panel (2018–2025), multiple urban industries, vacancy postings + firm adoption signals + city digital-infrastructure controls. Sample sizes in regressions ~48–52k city–occupation–week cells.
• Skill measurement: Instruction-tuned LLM extracts skills, tools and proficiency from vacancy text; mapped to O*NET/ESCO via string + embedding similarity; transformer embeddings used to form city–occupation–time skill vectors.
• Outcome measures: share of complementary skills, share of routine skills, human–AI collaboration index, cross-functional coordination index, and embedding-based semantic migration (Euclidean distance between successive period skill vectors).
• Causal identification:
  • Staggered difference-in-differences with event-time coefficients; treated event = city entering top quartile of GenAI diffusion.
  • Shift–share instrument: baseline occupational shares × exogenous foundation-model capability upgrades used in first stage; IV/2SLS and robustness checks (synthetic control, weighted regressions) reported.
• Quantitative effects (representative estimates):
  • Event-study: pre-event coefficients ≈ 0; τ=0: complementary skill share +0.50 percentage points (p<0.10). τ=1..4: +1.40, +2.60, +3.20, +3.80 p.p. (τ≥1 estimates p<0.01). These are ~0.18–0.49 standard deviations.
  • OLS regression: 1-unit increase in standardized GenAI diffusion → complementary skills +0.95 p.p. (SE 0.18), routine skills −0.68 p.p. (SE 0.15). Human–AI collaboration +0.41 (SE 0.07); cross-functional coordination +0.36 (SE 0.09).
  • Mechanism controls (human–AI and coordination) reduce event-time coefficients by ~25% on average but do not fully explain the effect.
• Robustness & reproducibility: pipeline containerized; model versions, prompts, and ontology-mapping logged; gold-standard labeled set used to compute precision/recall/F1 for extraction and alignment.

Data & Methods

• Data sources: scraped job boards, corporate career pages, headhunter sites; listed-company disclosures and developer docs for firm adoption signals; city-level infrastructure from yearbooks/open data.
• Text processing: HTML/URL stripping, de-duplication, tokenization, language detection; instruction-tuned LLM extracts skill tokens + proficiency; ontology alignment via string matching + embedding similarity.
• Skill-vector construction: share-weighted average of skill-token embeddings per city–occupation–time; semantic migration = L2 distance between successive vectors.
• Causal econometrics:
  • Primary: staggered DID / event-study with city–occupation and time fixed effects, controls, and pre-trend checks.
  • Endogeneity check: shift–share IV predicting city–time GenAI exposure; IV-DID / 2SLS used to recover event-time effects.
  • Additional robustness: synthetic control and weighted regressions to account for heterogeneous timing/intensity.
• Validation: gold-standard labels to compute precision, recall, and F1 for skill extraction and ontology alignment; stability checked across years and platforms.

Implications for AI Economics

• Mechanism-focused policy: Evidence supports targeting reskilling/upskilling at specific skill dimensions (collaboration, coordination, cognitive complementarities) rather than treating whole occupations as vulnerable. Training programs should prioritize teamwork, workflow orchestration, and human–AI interaction skills.
• Firm strategy and task design: Firms are more likely to embed GenAI into coordination and collaborative workflows than to fully replace frontline roles. Human–AI task allocation and team-structure redesign are consequential levers for productivity and employment composition.
• Urban heterogeneity matters: High human-capital density and digital infrastructure accelerate complementary-skill gains; local industrial structure and institutional factors shape adoption and reallocation patterns. Regional policy responses should therefore be tailored to city-specific labor-market structures.
• Measurement & research agenda: The paper demonstrates a scalable pipeline (LLM extraction + ontology alignment + embeddings) for measuring within-job skill shifts. Future work should (a) refine extraction to reduce LLM noise/hallucination, (b) link vacancy-based skill measures to realized employment/wage outcomes, and (c) extend to non-urban or informal labor markets.
• Cautions for interpretation: Results rely on vacancy text as a signal of demand (not necessarily realized hires), LLM-based extraction has potential bias/hallucination despite validation, and residual endogeneity or unobserved channels may remain despite shift–share IV and robustness checks.

Concise takeaway: Generative AI diffusion in cities causally reallocates task bundles inside jobs toward collaborative, coordination-intensive, and human–AI complementary skills—policy and firm responses should focus on targeted reskilling and task redesign rather than occupation-level protectionism.