AI‑generated text feeds back into the training corpora of subsequent models, and recursive training drives model collapse. Recent work extends this to many models feeding each other, but most studies assume an even market split, whereas real generative AI is an oligopoly.
We evaluate two concerns in controlled ecosystems: (1) whether fewer, more uniform sources accelerate collapse, and (2) whether later models are pulled toward the oligarch’s output. The experiment uses 13 open‑source models of 1–4 B parameters, forming natural ecosystems of 3 to 13 players, plus an injected probe that pushes the top share to 90%. Each generation, every model’s output is mixed into a shared pool according to market share, and each model is retrained on that pool from clean base weights for five generations.
Within the tested range, neither concern materializes. Unequal splits barely affect the speed of collapse; the destination moves even less: varying share and identity knobs shifts the five‑generation endpoints by only a few percent of the common drift. An extreme share combined with the strongest injected bias still does not guarantee steering, and the topic shifts it produces leave only a faint trace on the ruler measuring collapse.
What sets the speed is who supplies the pool and how readily those suppliers are carried along: with every share held fixed, swapping the members of a K=3 ecosystem changes five‑generation drift by 2.8×; a share‑weighted index of each member’s susceptibility explains the speed differences across nineteen arms with $R^2 = 0.68$; and replacing half the pool with human text roughly halves drift without changing its course.
In summary, within the explored range concentration neither determines the destination nor the pace of collapse; the pace follows whose text fills the pool.
Review