Alexandra Wölfle, Desirée Schmuck
Generative artificial intelligence (GenAI) is increasingly shaping our everyday lives. While the new technology brings undeniable benefits, it also bears considerable risks, such as biases in the models. Drawing on the Stereotype Content Model (SCM), we thus explore gender stereotypes of warmth and competence in AI-generated ratings, stories, and images in comparison to human-generated content. We conducted a set of preregistered, automated content analyses of ratings and stories created by ChatGPT-4o-mini ( N = 999 each) as well as images created by DALL·E 3 ( N = 990). We further collected rating and story data through a quota-based survey-experiment among US adults ( N = 993) and created a corpus of Google Images ( N = 733) as a point of comparison. Stories were analyzed using a dictionary approach, while warmth and competence in images were operationalized through smiling and face-pitching and automatically assessed using AWS Rekognition . Findings showed that women are portrayed as warmer and less competent than men in ChatGPT ratings and DALL·E images, and that this difference is greater than in human-generated content. Women were again described as warmer, but also more competent than men in ChatGPT's stories, with no significant difference from human-generated stories. We also assessed stereotypes of warmth and competence ascribed to non-binary people with mixed results depending on the source and content type. Showing that gender-stereotypical differences are not only reproduced in AI-generated ratings and images but also exceed those observed in human-generated outputs, our findings have crucial implications for policymakers, developers, and users of GenAI alike.