Algorithmic reproduction of occupational stigma: visual prototype lock-in in Chinese text-to-image AI models

This article examines how text-to-image AI models reproduce occupational stigma through visual sign systems. Analyzing 608 images generated by four Chinese AI platforms from prompts relating to chengguan (urban management enforcement officers), the study reveals pronounced semiotic convergence: regardless of prompt variation, generated images default to stereotypical configurations of dark blue uniforms, middle-aged males, and serious expressions. The article identifies a Visual Prototype Lock-in mechanism whereby algorithms solidify statistical correlations into generative paths resistant to modification, transforming stigma reproduction from discursive construction to computational generation. Negative prompts reinforce rather than disrupt gender stereotypes. Cross-platform comparison reveals divergent strategies for processing stigmatized content, including temporal displacement and cultural othering. By introducing occupational identity as an independent dimension of algorithmic bias, the study demonstrates how generative AI endows bias with technological materiality and underscores the need for culturally representative training datasets.