AI Art's Hidden Bias: Stable Diffusion Pushes Artists Toward Stereotypes

You type "a scientist" into an AI image generator, and it gives you a man in a lab coat. Type "a nurse"—you get a woman. That's not you being creative; that's the AI's hidden bias at work. New research from Michelle Dutoit and Baptiste Caramiaux looked at this problem from the artist's perspective, with 14 creative professionals using Stable Diffusion (a popular AI-to-image tool).

The study asked each artist to create illustrations from two prompts: one very specific, one more open-ended. The result? Regardless of the prompt, artists experienced what the researchers call "normative behavior"—the AI's tendency to stick to invisible patterns and defaults. This showed up as repeating obvious templates, producing stereotypical output, and even adding or removing details the artist never requested. For example, an artist might ask for a "modern city" and the AI insistently adds recognizable landmarks, or drop a person from a scene without explanation.

Most importantly, these AI quirks made artists feel disempowered. Instead of being a helpful collaborator, the tool seemed to fight their creative choices. Artists described feeling "creative compromise"—having to accept the AI's baked-in style or spend extra time fighting it. The researchers suggest that if AI art tools made their decision-making more transparent (like showing why it chose certain elements), artists could take back control.

This is early research, with only 14 participants, and it's focused on one tool. But it points toward a bigger issue: as AI becomes a studio partner for millions of people, we need to understand how it shapes, and sometimes limits, what we create. The paper is part of a workshop on explainable AI in the arts, and it argues for giving artists more visibility into how these systems "think"—not just to fix errors, but to preserve genuine human creativity.