Dynamics of collective creativity in AI art competitions

Creativity is a fundamental part of cultural evolution, but the mechanisms by which groups produce novelty are notoriously difficult to infer from the historical record. Iterated learning experiments have shown that cultural transmission reliably distorts artifacts toward the inductive biases of learners, but most of this work uses linear chains of human participants, and it is unclear how these dynamics play out in the networked, human-AI systems that increasingly shape cultural production. In this study, we analyze one such system: Artbreeder, which hosts daily “remix parties” in which users iteratively build on each other's work from a single seed image, producing branching lineages of images co-created by humans and generative AI. Using 130,882 images from 368 remix parties over 13 months, we find that images become less novel over the course of a party and that each remix step moves less far from its parents, consistent with the stabilization seen in iterated learning experiments, and that late-stage images drift toward the platform's dominant portrait style. More novel “parent” images produce more novel and more complex “children.” More novel children attract more likes but are not remixed more, whereas more complex children are remixed less. Finally, larger remix parties produce images that are more novel but less complex.