Introduction
While much of the discourse surrounding artificial intelligence (AI) bias focuses on allocational harm—such as biased hiring algorithms—this research investigates a more subtle form of representational harm: lexical choice. Even when women are clearly depicted participating in sports, humans and the AI models trained on their data are significantly less likely to describe them by their athletic role (e.g., "surfer") compared to men. This "naming bias" sidelines women’s professional identities in favor of generic gender labels.
Quantifying the Naming Gap
The researchers analyzed the ManyNames dataset, which contains multiple human-provided labels for objects in images. They categorized these names into two groups: Taxonomic (generic labels like "woman" or "girl") and Sports-related (role-specific labels like "skateboarder" or "tennis player").
The results revealed a significant "naming gap." On average, 46% of names provided for men were sports-related, compared to only 35% for women. Logistic regression confirmed that the gender of the person in the image is a highly significant predictor ($p < 0.001$) of whether a sports-related name is chosen over a taxonomic one.
Taxonomic Defaulting and AI Propagation
This phenomenon, termed "taxonomic defaulting," suggests that speakers conceptualize women in sports primarily through their gender rather than their actions. For example, a female athlete on a skateboard is frequently labeled as a "girl" by the majority of annotators, whereas a male athlete in a nearly identical context is labeled a "skater."
Crucially, AI models do not just learn to identify objects; they inherit the societal lens of their annotators. A ResNet101-based model trained on this data successfully reproduced the human bias ($p < 0.001$), assigning significantly higher weights to sports-related names for images of men than for images of women. This propagation of bias reinforces the stereotype that sports are a masculine domain, potentially making it harder for future computer vision systems to accurately recognize and validate women's participation in these activities.
Conclusion and Future Work
The study highlights that representational harm is deeply embedded in the datasets used to train modern L&V models. While the current study was limited by a Western-centric dataset and a binary operationalization of gender, it opens the door for investigating how these biases are amplified in larger multimodal models like CLIP. Addressing this naming gap is essential for creating AI that accurately reflects the diverse roles individuals play in society.
Summary updated