Can Prompt Modifiers Control Bias? A Comparative Analysis of Text-to-Image Generative Models

Prompt modifiers have limited effectiveness: While adding words like "diverse" or "inclusive" to prompts can influence outputs, the study reveals this approach is not a reliable or systematic solution for controlling bias in text-to-image models like Stable Diffusion, DALL-E 3, and Adobe Firefly.

Biases persist across multiple dimensions: The research documents that these generative models encode societal biases not just around gender, but also across race, geography, and culture—suggesting the problem is widespread and multifaceted rather than isolated to single categories.

Prompt order matters: The sequencing of words in prompts affects how biases appear in generated images, indicating that users cannot simply add bias-correction terms anywhere and expect consistent results.

No industry standard exists yet: The paper emphasizes there is currently no uniform or agreed-upon standard for measuring or controlling bias across AI image generators, making it difficult for developers and users to compare safety across different models.

New framework needed for accountability: The authors propose a "bias sensitivity taxonomy" and call for common metrics to evaluate how AI models exhibit and respond to bias-correction requests, suggesting that future progress requires industry-wide standardization rather than individual company solutions.