Model Concentration Barely Changes Collapse in Multi-Model Ecosystems

Summary

AI-generated text is increasingly fed back into the training data of later models, raising concerns that recursive training may cause model collapse. This study tests whether concentration in a multi-model ecosystem makes collapse faster or steers models toward the dominant supplier’s output. The authors use 13 open 1B–4B models in ecosystems containing three to 13 participants, plus an injected probe that raises one share to 90%. At each generation, outputs are mixed into a shared pool according to market share, and every model is retrained from clean base weights for five generations. Within the tested range, unequal shares barely change either the speed of collapse or the final destination; even the strongest injected bias does not reliably steer the ecosystem. The identity of the suppliers matters more: replacing the members of a three-model ecosystem changes five-generation drift by 2.8x, while a share-weighted susceptibility index explains differences across 19 experimental arms with R² = 0.68. Replacing half the synthetic pool with human text roughly halves drift without changing its direction. The authors conclude that concentration alone neither determines the pace nor the endpoint of collapse; the composition and susceptibility of the text suppliers are more important.