Title: The Stochastic Parrot in the Coal Mine: Model Collapse Poses a Threat to Low-Resource Communities

Abstract:

As artificially generated content becomes increasingly prevalent, the issue of model collapse—the decline in performance that occurs when generative models are trained on data produced by earlier models—has emerged as a growing concern. Previous critiques of large language models have pointed to their propensity to replicate common patterns found in training data, their dependence on massive datasets, and their significant environmental footprint. These elements collectively drive data degradation, amplify cultural biases, and result in inefficient resource utilization.

In this position paper, we synthesize these perspectives to argue that model collapse undermines current initiatives aimed at democratizing artificial intelligence. By diminishing training efficiency and distorting data distributions away from the extremes of their support, model collapse disproportionately harms marginalized and low-resource communities. We explore both the cultural and environmental ramifications of this phenomenon, contextualize our stance within recent scholarly discussions on model collapse, and conclude with a call to action. Additionally, we propose preliminary strategies for mitigating these adverse effects.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC