Occupational gender stereotypes are widely embedded in social cognition and increasingly reproduced through generative artificial intelligence (AI). Two mainstream Chinese generative AI models (DeepSeek V3 and Qwen 2.5) were audited by eliciting occupation–gender pronoun associations for 72 census-anchored occupations using a standardized questionnaire and an automated testing pipeline. Each occupation was queried in 1000 independent rounds, yielding 2,880,000 item-level observations. The results show that, for both models, the fitted relationship between census female shares and model-implied female pronoun associations follows an S-shaped pattern. This pattern is consistent with a dominance-amplifying mapping that pushes male-dominated occupations toward lower female attribution and female-dominated occupations toward higher female attribution. Meanwhile, women’s overall visibility is consistently shifted downward: when the census benchmark is 50% female, the predicted female proportion remains below parity at 48% in DeepSeek and 43% in Qwen. Cross-model comparisons reveal substantial heterogeneity in bias profiles: DeepSeek primarily compresses female attribution in male-dominated occupations, whereas Qwen amplifies female dominance in occupations where women already predominate. Overall, these findings characterize a multi-layered output-level bias pattern combining structural amplification with a system-wide downward shift in women’s aggregate visibility.
Liu et al. (Mon,) studied this question.