DALL-E Mini's Unexplained Obsession With Generating Women in Saris
Users discover that running blank prompts in DALL-E mini overwhelmingly produces images of brown-skinned women in saris. AI researchers point to poor tagging and incomplete training datasets as the likely cause of this strange glitch.
Users of the viral AI image generator DALL-E mini discover a bizarre pattern when they submit completely blank prompts. Instead of producing random or error messages, the system consistently generates portraits of brown-skinned women wearing South Asian saris. One Brazilian screenwriter tests this phenomenon extensively, running the blank command thousands of times over a 10-hour period to build a repository of over 5,000 unique images that mostly feature this specific subject.
The mysterious output confuses even the creator of the open-source AI model, who bases the project on OpenAI's highly restricted DALL-E 2. While the original system produces hyper-realistic art for a select group of researchers, DALL-E mini offers the public a stripped-down, accessible version. However, this wider access brings unexpected quirks to light, revealing that the AI relies heavily on specific visual patterns when it lacks explicit text instructions.
AI researchers explain that this strange fixation stems from flaws in the system's underlying training data. Shoddy tagging and incomplete datasets cause the algorithm to default to images of women in saris when it has no other contextual clues. This glitch highlights the broader issue of dataset bias, showing how AI models reflect the inconsistencies and gaps in the massive collections of images used to train them.