TLDR: A study by researchers from Togo AI Labs and Vizuara AI Labs reveals significant gender biases in Vision-Language Models (VLMs). By analyzing how VLMs associate face images with occupation and activity descriptions, the study found consistent male- or female-leaning associations across different labor categories and specific statements, with transformer-based models exhibiting slightly stronger biases. The research introduces a robust framework for measuring and understanding these embedded social stereotypes.
Vision-Language Models (VLMs) are powerful AI systems that can understand and connect both images and text, enabling capabilities like open-vocabulary recognition and zero-shot learning. However, new research highlights a critical concern: these models can inadvertently learn and amplify social stereotypes, particularly gender biases, from the vast amounts of web data they are trained on.
A recent study, titled “Vision-Language Models display a strong gender bias,” conducted by researchers from Togo AI Labs and Vizuara AI Labs, delves into this issue. The authors, Aiswarya Konavoor, Raj Abhijit Dandekar, Rajat Dandekar, and Sreedath Panat, investigated whether contrastive vision-language encoders exhibit gender-linked associations when pairing face images with phrases describing occupations and activities.
How the Study Was Conducted
The researchers assembled a dataset comprising 220 face photographs, evenly split by perceived binary gender, and 150 unique statements. These statements were categorized into six types of labor: emotional, cognitive, domestic, technical, professional roles, and physical labor. For each statement, they calculated an “association score.” This score was determined by measuring the difference in how strongly the VLM associated the statement with male faces versus female faces. A positive score indicated a stronger association with male faces, while a negative score indicated a stronger association with female faces.
To ensure the robustness of their findings, the team used bootstrap confidence intervals and a “label-swap null model.” The latter helped estimate the level of association expected if no real gender structure were present, providing a baseline to distinguish genuine biases from random noise. The study evaluated several pre-trained CLIP-style dual encoders, including transformer-based models like ViT-B/32 and ViT-L/14, and ResNet-based models like RN50 and RN101.
Key Findings: Consistent Gender Biases
The study revealed consistent gender biases across all evaluated models, with transformer-based architectures generally showing slightly stronger magnitudes of bias. All four models exhibited biases significantly greater than what would be expected by chance, with the ViT-B/32 model showing the highest ratio of observed bias to null model bias (2.00).
When examining specific labor categories, clear patterns emerged:
- Female-leaning associations: Emotional labor, cognitive labor, and technical labor.
- Male-leaning associations: Domestic labor, professional roles, and physical labor.
For instance, emotional labor (-0.178), cognitive labor (-0.410), and technical labor (-0.898) showed strong female associations, while domestic labor (+1.180), professional roles (+0.835), and physical labor (+0.297) tended to be male-associated. These patterns were consistent across different models, highlighting deeply embedded stereotypes.
The research also identified specific statements with strong biases. For example, the ViT-B/32 model associated males more with terms like “firefighter,” “carpenter,” and “truck driver,” and females more with “nurse,” “teacher,” and “caregiver.” Similarly, other models showed associations like “pilot,” “CEO,” and “engineer” with males, and “therapist,” “counselor,” and “librarian” with females. These findings align with real-world occupational gender imbalances, indicating that VLMs are reflecting and potentially amplifying existing societal biases.
Also Read:
- AI’s Hidden Flaws: Uncovering Cognitive Biases in General-Purpose AI for Software Engineering
- Unpacking Prompt Sensitivity: A Deep Dive into LLM Robustness
Implications for AI Development
The findings underscore that architectural choices and the data used for pretraining significantly influence the direction and magnitude of learned associations in VLMs. This research provides a transparent and reproducible method to measure gender-linked associations, offering a valuable framework for practitioners to understand how an encoder’s learned geometry relates to socially relevant categories. Addressing these biases is crucial for developing more equitable and fair AI systems. You can read the full research paper here: Vision-Language Models display a strong gender bias.
