This project focuses on open-source models because open models allow for more transparent interrogation of the full modeling pipeline, from training data sources and filtering methods to model architecture and prompt handling. Closed models are by no means exempt from bias, but the transparency offered by open-source models enables more replicable, scalable research into how specific choices in dataset construction or fine-tuning practices can influence representational outcomes.

The ease of forking, fine-tuning, and deploying open models means that their outputs are already influencing countless downstream applications, and often beyond the oversight of any single institution.  Platforms like Hugging Face, GitHub, and Kaggle have democratized access to these technologies, while organizations like LAION have contributed massive open datasets that power many popular generative models. These platforms serve as both infrastructure providers and de facto governance bodies shaping how models are shared, documented, and deployed.

By contrast, closed models like Midjourney or DALL·E 3 restrict access to their training data, internal weighting, and even their prompt parsing mechanisms, making it difficult for us to systematically trace or audit the origins of their biases. Both types of systems are susceptible to harmful patterns, but open-source models facilitate efforts to audit the technology, document bias, and meaningfully mitigate it.

If we're moving into a future where the images these models output are going to be used to represent us and illustrate our world, understanding the biases embedded in the technology becomes a matter of public infrastructure and digital equity.

Why Open-source?