Unveiling Hidden Biases: Analyzing Professional and Gender Perception Patterns in Large Language Models

This study explores how the Claude and Gemini language models exhibit gender bias in relation to occupational roles within Turkish texts. To evaluate the extent of bias in each model, data from 22 different occupational groups was analyzed, focusing on the gender distributions of names generated for each occupation. The findings indicate that the Claude model shows a stronger tendency to predict male gender for traditionally male-dominated professions such as engineering and military service, while it more frequently associates female gender with occupations like nursing and teaching. By comparison, the Gemini model demonstrates a more pronounced male bias in creative fields, particularly in artistry and writing. Additionally, the study examined the models’ gender prediction performance in Turkish texts as a sub-question, investigating whether these predictions exhibit systematic biases. The results suggest that the gender predictions made by these models reflect societal biases, highlighting the necessity for bias mitigation strategies in datasets to ensure the ethical application of language models.