Text-to-image models are everywhere now, with thousands uploaded to Hugging Face alone. But the biases baked into these systems are just as widespread—and they’re not accidental. They’re structural, appearing at nearly every stage of the pipeline, from the training data itself to the filters applied after an image is generated.
The most obvious culprit is the data. Massive multimodal datasets like LAION-5B, MS-COCO, and VQA v2.0 have been repeatedly shown to contain harmful associations and skewed representations. If most of the training text is in English, you get Western values as the default. Hugging Face’s own Stable Bias project found that image generations lack diversity and perpetuate stereotypes about cultures and identity groups. When you ask DALL-E 2 for a CEO, you get a white man. Ask for a manager, same thing. The pattern repeats across models.
But the data isn’t the only problem. The CLIP model that guides text encoding in Stable Diffusion and DALL-E 2 has well-documented biases around age, gender, and race—treating white, middle-aged, and male as the default. Even filtering steps meant to clean up training data can backfire. The DALL-E 2 team found that filtering actually amplified certain biases, possibly because the original data already represented women in more sexualized contexts or because the filters themselves carried baked-in assumptions. Post-hoc safety filters are equally shaky, with red-team efforts showing they’re not robust to different kinds of problematic content.
What’s striking is how persistent this problem remains even as the technology races forward. Tools like Firefly and Shutterstock are already deployed in commercial contexts, which means these biases aren’t academic—they’re shaping what millions of people see when they search for stock imagery or generate marketing assets. The relationship between bias and values is intimate, and trying to encode the evolving values of a population into a single static model may simply not be feasible. Some researchers are exploring the latent space of diffusion models to steer generations along axes like gender, but that work is still in its infancy. The uncomfortable truth is that every stage of these systems, from dataset creation to final output filtering, introduces its own flavor of distortion. Fixing one stage won’t fix the whole. The question isn’t whether bias exists—it’s whether anyone building these systems is willing to treat it as a design constraint rather than an afterthought.
