Chinese listed firms that adopt AI expand and future-proof the skills they advertise: adopters both specify current occupation requirements more precisely and increasingly demand skills that only later became formalized, suggesting AI lets firms lead changes in occupational standards.

Main Finding

Firm-level AI capability causally increases the richness and forward-looking nature of skill demands in online job vacancies. AI adoption leads firms to (i) specify more occupation-aligned skills with greater textual clarity (less vague language, more consistency) and (ii) list more non-aligned, “forward-looking” skills that predate formal updates to occupational taxonomies—evidence that AI both stabilizes current signaling and enables firms to lead changes in skill standards.

Key Points

• Data: 14 million Chinese online job vacancies (2018–2022) from major platforms (Zhaopin, 51job, Liepin), matched to listed firms and subsidiaries.
• AI capability measure: firm AI patent stock (Incopat), perpetual inventory method (15% depreciation), AI patents identified by IPC codes for ML, neural nets, CV, NLP.
• Main empirical result: within-firm increases in AI capability are associated with higher counts of (a) occupation-aligned skills and (b) non-aligned skills in job postings; IV estimates using patent-examiner leniency confirm causality.
• Mechanisms:
  • Information/clarity channel: AI-capable firms produce more consistent skill descriptions and fewer ambiguous expressions.
  • Anticipation channel: AI-capable firms list skills that were not in 2018 taxonomies but were later codified in 2022 (termed forward-looking skills), suggesting predictive identification of emerging skill needs.
• Methodological contribution: a scalable NLP pipeline mapping Chinese vacancy text to ESCO’s 13,939-skill taxonomy using an Extreme Multi-Label Classification (XMLC) bi-encoder trained via contrastive learning, with LLM-generated synthetic training pairs to overcome labeled-data scarcity.
• Robustness: effects robust to firm, occupation, year fixed effects, alternative AI-stock depreciation rates, different similarity thresholds for mapping, and various specification checks probing the IV exclusion restriction.

Data & Methods

• Core datasets:
  • 14M job postings (Zhaopin, 51job, Liepin), 2018–2022.
  • AI patent records from Incopat (mapped by IPC to ML/NLP/CV categories).
  • Firm financials and controls from CSMAR; firm–occupation–year panel.
• Skill extraction:
  • Target taxonomy: ESCO (13,939 skills).
  • Model: bi-encoder XMLC embedding vacancy sentences and ESCO labels into shared semantic space; trained with contrastive learning.
  • Data augmentation: LLMs generate synthetic sentence–skill pairs to create large-scale labeled training data capturing implicit expressions.
  • Occupation assignment: title similarity mapping to ONET occupations; ONET task descriptions semantically mapped to ESCO to define occupation-aligned skill baselines (2018 benchmark).
• Identification:
  • Baseline regressions with firm × occupation × year panel and fixed effects to use within-firm variation.
  • Instrumental variable: patent-examiner leniency at CNIPA (historical grant rates of assigned examiners) that affects probability of AI patents being granted → exogenous variation in AI patent stock.
  • Tests: textual clarity measures (within-cell consistency; frequency of vague phrases), forward-looking skill tests by comparing 2018 vs 2022 taxonomy alignment.
• Sensitivity and robustness analyses: alternative depreciation, mapping thresholds, placebo checks on instrument, correlations with non-patent outcomes.

Implications for AI Economics

• Signaling and matching: AI can improve employer-side signaling by clarifying existing requirements and surfacing emerging skill needs—potentially improving match quality and reducing search frictions. This complements literature showing AI may degrade worker-side signals (e.g., AI-assisted cover letters), highlighting that AI’s net effect on matching depends on how it’s deployed across market sides.
• Skill dynamics and labor demand: AI adoption expands firms’ articulated skill portfolios, accelerating demand for both current and novel skills. This supports the view that technological capability drives endogenous upskilling and can lead firms to proactively reshape job content ahead of formal occupational-standard revisions.
• Occupational standards & measurement: Firms with AI capability may lead taxonomy evolution; researchers and policymakers relying on static taxonomies risk underestimating emerging skill demands. The paper also provides a scalable measurement approach (XMLC + LLM augmentation) that can improve monitoring of skill change in real time.
• Policy and training: Faster emergence of forward-looking skill demands implies urgency for adaptive retraining and continuous education. Labor-market policies should account for asymmetric firm capabilities: firms with fewer AI resources may lag in identifying necessary reskilling needs for incumbent workers.
• Industrial and regulatory considerations: Patent-based AI capability measures matter—firms investing in AI R&D are more likely to alter hiring practices and skill signals. Policymakers monitoring AI diffusion should consider how AI adoption changes labor demand composition even in roles not directly automated.
• Research directions: Assess downstream effects on hiring outcomes, match quality, wages, and worker upskilling; generalize beyond listed firms and China; evaluate potential complementarities and conflicts between AI-driven employer-side clarification and worker-side signal dilution; refine instruments for causal inference on other AI channels.

Limitations to keep in mind: patent stock is an imperfect proxy (misses non-patented AI adoption); instrument validity rests on exclusion assumptions (examiner leniency only affects skills via granted patents); LLM-generated training data may introduce semantic bias; sample focuses on Chinese listed firms and online vacancies, which may limit generalizability.

Overall, the paper shows that AI is not only an automation technology but also a powerful information technology that reshapes employer signaling and the evolution of skill demands—both stabilizing current recruitment information and accelerating the market’s anticipation of future skills.