The meta-model of AI image generation seems to have difficulty dealing with the simple reality of cross-cultural relationships; When testing the model from The Verge website, the AI image generator from Meta showed a surprising inability to imagine images of Asian individuals alongside their Caucasian counterparts.
Despite numerous attempts using various directives such as “Asian man and Caucasian friend,” “Asian man and white wife,” or “Asian woman and Caucasian husband,” the tool mostly generated images showing only two Asian individuals.
This flaw extends to depictions of mixed-race couples and cross-race friendships, as the AI model constantly returns images that do not reflect the racial mix specified for it in the written directions.
For example, the prompt “Asian man and white woman smiling with dog” resulted in images of only two Asian people, even after adjusting the written description.
The site also noticed similar problems with the guidelines requesting depictions of friendships between different races, as the tool was again tending to display images of Asian individuals.
What is striking is that when the site moved to focus on photos of South Asian individuals, the results were slightly more accurate but still varied. The tool was able to produce a photo of a South Asian man with a Caucasian woman on one occasion, but then returned to producing photos of South Asian individuals only.
Beyond these obvious biases, the AI model results also point to deeper, more systematic biases.
Despite numerous attempts using various directives, the tool mostly generates images showing only two Asian individuals (The Verge)
For example, Gadget’s depiction of Asian women mostly aligns with East Asian features and lighter skin tones, ignoring the enormous diversity of Asian demographics. In addition, there is a clear tendency to use culturally specific clothing, such as the bindi and sari, which are famous for India, without any explicit instructions for the model to use this type of clothing.
These biases are not only related to racial fallacies, but also reflect stereotypical and age biases. The model often portrays the image of Asian women as young and East Asian in appearance, ignoring the age and ethnic diversity in various parts of the Asian continent.
The shortcomings and biases of the AI meta-model raise several important concerns about the inclusivity and cultural sensitivity of AI tools, and this situation underscores the need for AI development to be more reflective of global cultural diversity and avoid continuing to be influenced by the stereotypes and biases that dominate the world. .
The Verge’s experiment with a meta-model may be a cautionary example about the potential pitfalls in developing generative AI models, highlighting the importance of the different datasets on which these models are trained. As we move into the future, countering such biases will be essential to developing AI systems that truly understand and represent the rich texture of the human experience.
