I’m still trying to generate an AI Asian man and white woman

🌈 Abstract

The article discusses the challenges and inconsistencies faced by various AI image generators, such as Meta's, Google's Gemini, DALL-E, and Midjourney, in creating images depicting Asian men and white women. It highlights the race-swapping issues, the inability to handle certain prompts, and the tendency to rely on stereotypes in the generated images.

🙋 Q&A

[01] Challenges with AI image generators

1. What are the key issues the article highlights with AI image generators?

• The article discusses several issues with AI image generators, including:
  • Meta's AI image generator built into Instagram messaging failed to create images of an Asian man and white woman using general prompts, instead changing the woman's race to Asian
  • Meta appeared to have blocked prompts with keywords like "Asian man" or "African American man"
  • Other AI image generators like Google Gemini, DALL-E, and Midjourney also struggled with generating images of Asian men and white women, producing inconsistent and confusing results

2. How did the author investigate these issues?

• The author teamed up with The Verge's Emilia David to generate AI-created images of Asian people across multiple platforms, including Google Gemini, DALL-E, and Midjourney
• The author also tried various prompts on Meta's AI image generator built into Instagram messaging and documented the results

3. What were the key findings from the author's investigation?

• Google Gemini refused to generate Asian men, white women, or humans of any kind
• DALL-E and Midjourney struggled to generate images that accurately depicted an Asian man and a white woman, producing confusing and inconsistent results
• Meta's AI image generator had issues with race-swapping, changing the race of the woman in the prompts
• The most consistent way for the AI models to depict an Asian man and a white woman was by placing them in an "academic setting"

4. How did the companies respond to the author's findings?

• Google, OpenAI, and Midjourney did not respond to the author's requests for comment
• Meta spokesperson Tracy Clayton acknowledged the issues, stating that the technology is new and "won't always be perfect," and that they are working on improving their models

[02] Implications and Biases

1. What does the author suggest about the biases and implications of these AI image generation issues?

• The author questions what kind of biases are "baked into training sets" that lead to these issues, and what it means that the most consistent way for the AI models to depict an Asian man and a white woman is in an "academic setting"
• The author also notes certain "ticks" or patterns in the generated images, such as women of all races wearing the same white floral dress, the presence of flowers surrounding couples, and the lack of diversity in terms of age, body type, and hair color

2. How does the author feel about the companies' responses to these issues?

• The author expresses frustration with the "radio silence" from the companies or their use of "generalities" in responding to these issues, rather than providing a clear explanation of what is going wrong and how they plan to address it.