AI exposure in China has climbed rapidly and appears to displace traditional, routine occupations—workers in high-exposure roles earn less and vacancies take longer to fill—while generating little measurable new employment in emerging occupations.

Main Finding

AI exposure in China (measured at the task/occupation level using LLM- and Sentence‑BERT–based semantic similarity) has risen steadily from 2019–2024 and is associated with (i) net substitution in traditional occupations but no clear creation effect in emerging occupations, (ii) lower wages concentrated in routine‑intensive occupations (partly offset where task complexity is higher), and (iii) worsened matching outcomes — lower job filling rates and longer recruitment times — consistent with increased labor‑market friction.

Key Points

• Measurement innovation: an AI exposure index constructed by measuring semantic similarity between AI capabilities and occupational task requirements using large language models and Sentence‑BERT.
• Data source and period: retrieval‑augmented generation (RAG) applied to online job postings in China, 2019–2024.
• Exposure trend: AI exposure rises across tasks and across most occupations over the sample period.
• Substitution vs creation:
  • Clear substitution effects in traditional, routine occupations (reduced employment opportunities).
  • No significant creation effect detected for emerging/new occupations within the study window.
• Wages:
  • Occupations with higher AI exposure show lower wages on average.
  • The negative wage association is strongest where routine task share is high.
  • Task complexity mitigates the negative wage effect (i.e., complex tasks are less substitutable).
• Matching frictions:
  • Higher AI exposure correlates with lower job filling rates and longer time-to-hire, implying increased mismatch or recruiting difficulty in exposed occupations.
• Heterogeneity: effects differ by occupational task composition (routine vs non‑routine) and task complexity.

Data & Methods

• AI exposure index:
  • Constructed by computing semantic similarity between structured descriptions of AI capabilities and occupational task descriptions.
  • Models used: large language model to interpret capability descriptions plus Sentence‑BERT embeddings to compute similarity scores between AI features and task statements.
• Job-market data:
  • Large-scale scrape of online job postings in China (2019–2024).
  • Retrieval‑augmented generation (RAG) used to extract and standardize task descriptions, wage offers, posting duration, and filling status from unstructured listings.
• Empirical strategy:
  • Panel regressions at the occupation (or occupation-region/time) level linking AI exposure to outcome variables: number of job postings (employment opportunities), offered wages, job filling rates, and recruitment time.
  • Controls and specification elements (typical): time and occupation fixed effects, controls for local labor demand/supply conditions, occupation task composition (routine share, complexity), and robustness checks across alternative exposure specifications and subsamples.
• Identification caveats:
  • Observational associations; potential for omitted variables and reverse causality (e.g., declining occupations attracting more automation investment).
  • The index is based on semantic matches (potential measurement error) but benefits from locally sourced job text and modern embedding/LLM methods to better capture China‑specific tasks.

Implications for AI Economics

• Measurement: semantic similarity methods with LLMs and SBERT + RAG provide a scalable, locally relevant way to map AI capabilities to occupational tasks — improving risk assessments beyond imported indexes.
• Short‑run labor effects: evidence points to net substitution in exposed occupations rather than immediate net job creation, emphasizing short‑run displacement risks and the need to model heterogeneous task substitutability.
• Wage and inequality dynamics: AI exposure depresses wages more in routine occupations, suggesting potential for wage compression or increased inequality unless offset by upskilling or occupational mobility.
• Matching and frictions: longer hiring times and lower fill rates imply that automation can increase labor‑market frictions (skills mismatch, uncertainty about job content), suggesting complementarities between automation and active labor‑market policies (retraining, signaling, job‑search assistance).
• Policy relevance: targeted reskilling, incentives for task reallocation toward complex activities, and policies that reduce matching frictions (credentialing, labor intermediation) could mitigate negative impacts.
• Research directions: causal identification (natural experiments, instrumenting AI exposure), firm‑level and worker‑level microdata to trace transitions, long‑run creation effects, spatial heterogeneity, and evaluation of policy responses.