Position: the Stochastic Parrot in the Coal Mine. Model Collapse is a Threat to Low-Resource Communities

Researchers argue that model collapse—performance degradation when training generative models on synthetic outputs from prior models—disproportionately harms low-resource communities by reducing training efficiency, concentrating data toward high-frequency patterns, and amplifying cultural biases while increasing computational costs.