On gender stereotypicality in nouns and adjectives: Comparing humans, large language models and text-to-image generators

Abstract

Both humans and large language models (LLMs) are known to exhibit effects of gender stereotypicality. We conducted a series of studies to systematically assess to what extent humans’ and LLMs’ interpretational patterns align, how different kinds of linguistic expressions (role nouns vs. adjectives) contribute, and to what extent these patterns extend to text-to-image models. Experiments 1 and 2 test how gender-biased role nouns (e.g. plumber, nurse) and adjectives (e.g. powerful, kind) influence humans’ and GPT-4o’s assumptions about gender in a fill-in-the-blank task. Experiment 3 tests how role nouns and adjectives influence images created by the image generator DALL-E 3 (a text-to-image model). Our results show that humans, LLMs and text-to-image models’ outputs are all influenced by gender stereotypes but diverge in unexpected ways.