A leading Chinese hiring LLM systematically scores applicants aged 35 and over—and women—lower than younger male counterparts, suggesting the tool amplifies existing labour-market biases; the study calls for targeted fairness audits and regulatory oversight.

Main Finding

DeepSeek, a leading Chinese large language model used for automated résumé screening, systematically disfavors female applicants and candidates aged 35 and older. Using a simulated CV–job matching audit (4,080 CV–job pairs), the authors find statistically and practically meaningful score disparities: females and 35+ candidates receive lower suitability scores than comparable peers, and these patterns persist across linear (OLS) and nonlinear (XGBoost + SHAP) analyses.

Key Points

• Scope and contribution
  • First empirical audit of a major Chinese LLM (DeepSeek) in recruitment settings, extending AI fairness evidence beyond Western contexts.
  • Combines interpretable OLS with machine‑learning (XGBoost) and SHAP explanations to reveal both linear and nonlinear bias patterns.
• Core quantitative results
  • Overall deviation from a qualifications-only fairness benchmark: MAE = 6.8 (p < .001).
  • MAE by group: male = 4.2 vs female = 9.4; under‑35 = 3.9 vs 35+ = 13.1.
  • Mean score patterns: males scored higher than females across all age brackets; gap peaks around age 45 (male 69.09 vs female 61.29). Scores peak at ages 25–30 and decline notably after 35.
• Interpretability
  • SHAP analysis exposed how demographic features (age, gender) and their interactions nonlinearly influenced predicted suitability scores beyond what OLS captured.

Data & Methods

• Data generation and experimental design
  • Job postings: 255 scraped listings from BOSS Zhipin, stratified across STEM and technical job types.
  • CVs: a standardized résumé template produced 16 synthetic candidate profiles per job (total 4,080 CV–job pairs). Controlled attributes: education, region (major Chinese metros), skill keywords, and years of experience (0–20, increasing with age).
  • Demographics: 8 age identities {18, 22, 25, 30, 35, 40, 45, 50}; gender balanced 50/50.
• LLM scoring procedure
  • Model: DeepSeek-reasoner (April 2025 API version).
  • Prompt: zero-shot, structured instruction in Mandarin asking for a single integer score (1–100) representing interview suitability.
  • Automation: API calls for all pairs (1.5 s between calls) to remove conversational memory effects.
• Fairness benchmark and evaluation
  • Constructed benchmark score per candidate based on qualifications (education, experience, skills); fairness deviation measured as MAE between DeepSeek score and benchmark.
• Estimation strategy
  • OLS regressions: score ~ age + gender (+ interaction age×gender) for transparent effect estimation.
  • XGBoost: modeled score with demographic and CV features to capture nonlinearities; data split 80/20 train/test.
  • Model tuning: grid search over 216 parameter combinations with 5‑fold CV (1,080 fits).
  • Explainability: SHAP used to quantify and visualize marginal contributions of features.

Implications for AI Economics

• Labor market fairness and access
  • Automated LLM screening can reproduce and amplify existing labor‑market discrimination (gender and age), potentially exacerbating inequality in hiring outcomes at scale.
  • The pronounced penalty for 35+ candidates implies automation may institutionalize informal age thresholds (e.g., the "35 barrier") into hiring pipelines.
• Firm-level decisions and efficiency trade-offs
  • While LLM screening increases throughput and reduces screening costs, it can produce efficiency gains at the expense of equity; firms seeking diverse talent may incur hidden selection costs if relying on biased LLM outputs.
  • Hidden bias can distort labor supply signals, affecting wage bargaining, career investment decisions, and human‑capital accumulation across cohorts and genders.
• Policy and regulation
  • Results support the need for regulatory auditability and transparency for proprietary LLMs used in high‑stakes decisions (hiring). Mandates for fairness testing, reporting, and human‑in‑the‑loop safeguards are warranted.
  • Benchmarking and external audits (including culturally specific audits) should be part of deployment requirements, because bias patterns can be context‑specific.
• Research and market design
  • Combining interpretable econometric methods with explainable ML is a productive approach for diagnosing algorithmic bias and informing corrective interventions (e.g., reweighting, counterfactual data augmentation, prompt engineering).
  • Economic models of hiring should incorporate algorithmic intermediaries as endogenous actors that can alter selection friction, match quality, and inequality dynamics.
• Practical interventions
  • Short term: apply human oversight, blind features, fairness-aware prompts, or score calibration to reduce demographic leakage.
  • Long term: require model provenance, disclosure of training data composition, and routine bias audits, especially in markets with strong preexisting discrimination.

Caveats and limitations (noted by authors) - Synthetic CV audit: controlled profiles enable causal inference on demographic effects but do not observe real hire outcomes (external validity to live hiring may be partial). - Scope: analysis focused on age and gender within Chinese STEM/technical jobs and on one proprietary model/version (DeepSeek‑reasoner, Apr 2025); results may vary with other models, prompts, or real-world deployments. - Opaque model internals: proprietary, black‑box LLMs limit understanding of training‑data origins and mechanisms producing bias.

Suggested next steps - Field experiments with real application pipelines and downstream hiring outcomes. - Intervention studies testing prompt‑based debiasing, data‑level corrections, and post‑processing fairness constraints. - Cross‑model and longitudinal audits to track whether biases persist across model updates and deployments.