research
The Biggest Model Barely Changed Where the Copies Ended Up
Across controlled ecosystems of up to thirteen models, unequal market share scarcely changed recursive collapse; supplier susceptibility changed its speed 2.8-fold.
Summary
Across controlled ecosystems of up to thirteen models, unequal market share scarcely changed recursive collapse; supplier susceptibility changed its speed 2.8-fold.
Researchers repeatedly retrained ecosystems of three to thirteen open 1–4B models on a shared pool containing their own generated text, varying market concentration through a 90% injected supplier and following five generations. Within that controlled range, greater concentration barely changed the speed or destination of model collapse. Holding shares fixed but changing which three models supplied the pool changed five-generation drift by 2.8×, and a share-weighted susceptibility index explained 68% of speed differences across nineteen arms. Replacing half the pool with human text roughly halved drift. The result is a bounded synthetic ecosystem study: it does not establish how proprietary production models or real web-scale corpora will evolve.
Why it matters
Across controlled ecosystems of up to thirteen models, unequal market share scarcely changed recursive collapse; supplier susceptibility changed its speed 2.8-fold.
Limits and context
- The result is a bounded synthetic ecosystem study: it does not establish how proprietary production models or real web-scale corpora will evolve.
Key claims
Across controlled ecosystems of up to thirteen models, unequal market share scarcely changed recursive collapse; supplier susceptibility changed its speed 2.8-fold.
Qualification: The result is a bounded synthetic ecosystem study: it does not establish how proprietary production models or real web-scale corpora will evolve.
Evidence: source-2026-09-13-002
Sources
- arXiv preprint 2609.11146arXiv · primary research
Corrections
No corrections have been recorded for this story.