Brief

A new study published on arXiv analyzed over 1.5 million occupational personas generated by four major large language models, including GPT-4 and Gemini 2.5. The research found that these models tend to create less diverse demographic representations compared to real-world data, often compressing occupations into a single dominant profile. The audit revealed consistent underrepresentation of White and Black workers, and overrepresentation of Hispanic and Asian workers, with biases amplifying existing occupational segregation and in some cases leading to near erasure of certain demographics. AI