Structure-Aware Diversity Pursuit

as an AI Safety Strategy against Homogenization

Ian Rios-Sialer

“Homogenization is an AI safety issue; diversity dynamics [are] at the core of the Alignment Problem.”

Generative AI models reproduce the biases in the training data and can further amplify them through mode collapse. We refer to the resulting harmful loss of diversity as homogenization. Our position is that homogenization should be a primary concern in AI safety. We introduce xeno-reproduction as the strategy that mitigates homogenization. For auto-regressive LLMs, we formalize xeno-reproduction as a structure-aware diversity pursuit. Our contribution is foundational, intended to open an essential line of research and invite collaboration to advance diversity.