Testing for the Role of Linguistic Distributional Knowledge In the Gender-Career Implicit Association Test: A Trial Level Analysis

We examined whether linguistic distributional representations derived from a range of language models can account for gender-career bias in the IAT. Using trial-level linear mixed-effects models (N = 42849, Observations: 4,721,084), we found that language models do capture behavioral patterns, and that classical models (e.g., word2vec) generally outperform LLMs.