The Male Abstraction Bias in Internet Culture and Large Language Models

arXiv:2611.05730v1 Announce Type: new Abstract: Two foundational aspects of human cognition, abstractness and gender, are typically examined as separable. Here, we find they are systematically linked throughout the entire space of concepts spanning single (N = 40,000) and multi-word expressions (N = 67,000): across billions of words and millions of images on the internet -- and across numerous measurement techniques -- more abstract concepts are more associated with men, an effect which replicates in languages around the world. This pattern not only holds across conceptual domains, but it also maps onto grammatical structure: verbs and adverbs are more abstract and more male-coded, a system-wide trace of stereotypes accentuating agentic qualities in men. Furthermore, this bias emerges in human and AI responses in free association tasks, indicating its internalization and its algorithmic reproduction. The male abstraction bias may underlie and unify a range of stereotypes, evidencing gender as a core organizing dimension in conceptual space.

Male abstraction bias found in internet culture and large language models, perpetuating stereotypes.

The finding that abstract concepts are systematically associated with men in language and AI suggests a reinforcement of gender biases and stereotypes via AI models and internet culture. This perpetuates gender inequality and biased social perceptions. The effect is broad and embedded in grammar and AI responses, increasing its societal impact. However, direct harm levels are indirect and gradual.

The study claims that AI, specifically Large Language Models, reproduce a 'male abstraction bias' found in internet data. This bias links abstract concepts more with men across languages and conceptual domains, potentially perpetuating gender stereotypes algorithmically.

The text describes a systematic bias in language models that reinforces gender stereotypes. This bias could exacerbate existing social inequalities and misrepresentations.

SOURCE     arXiv cs.CY (Computers and Society) (tier 1)
   ↓
DOCUMENT   9e9dd626-9f62-4508-afd3-0eba54509d5f
           https://arxiv.org/abs/2611.05730
   ↓
EVIDENCE   2 extracted excerpts
   ↓
MODEL RUN  5 runs, methodology 1.0.1
   ↓
SCORE      -55  (Adverse)
   ↓
CONFIDENCE 62%