Gender Bias in Automated CV Evaluation: Evidence from Counterfactual Simulations Using Synthetic Data from Mexico

This paper investigates the presence of gender bias in automated CV evaluations conducted by a large language model (LLM). Using over 14,000 synthetically generated CVs representative of the Mexican labor market across six occupational categories, we implement a counterfactual design that isolates the effect of perceived gender by switching only the name and reported gender of each candidate. The analysis reveals systematic and occupation-specific biases: female candidates receive higher scores when presented as male in traditionally masculine roles (e.g., truck driver), while male candidates gain when reclassified as female in feminized occupations (e.g., nursing, elementary teaching). Notably, we document a statistically significant and operationally meaningful pro-female bias in the high-status Chief Financial Officer role. These asymmetries persist under deterministic prompting (temperature = 0), ruling out randomness as a confounder. Our design is contextually grounded, using names, educational institutions, and employers common in Mexico, and offers a scalable methodology for bias auditing in LLMs. The findings highlight the necessity of localized fairness assessments and raise concerns about the equity implications of deploying general-purpose AI tools in personnel selection.