Algorithmic reproduction of social inequality: language attitude bias in Chinese and English pre-trained language models

Pre-trained language models (PLMs) are used in education, hiring, and moderation, raising the question of whether these models could replicate social hierarchies and symbolic power structures. This study employs the Fill-Mask Association Test (FMAT) in four principal Chinese PLMs with reference to the English PLMs. Language attitude bias is systematically characterized by gender, occupation, and age. The results show that Chinese PLMs display stronger male-oriented bias in gender evaluations; all models consistently favor high-status professions in occupational bias; and in the age dimension, both Chinese and English PLMs prefer younger groups, though the Chinese models show greater internal variability. Status also emerges as the most stable and directionally consistent bias across models. These results illustrate how algorithmic discourse may mimic and sustain social ideologies, making the case for critical scrutiny and diverse data governance as a means to realize equity in digital communication.