Training AI image generators on their own past output doesn't always end the same way.
Researchers studying diffusion models, the generative systems behind many AI image tools, tested what happens when a model trains on data produced by earlier versions of itself. Prior studies disagreed on the outcome: one line of work found that fully replacing real training data with synthetic data wrecks a model within a few generations, while another found that simply accumulating real data alongside synthetic data prevents the problem entirely. This paper tests the regime most real-world pipelines actually use: keep every old dataset, but train each new model on a fixed-size sample pulled from that ever-growing pool, so the real-data fraction thins out over time without anyone deleting anything. Across a synthetic spiral dataset and the MNIST, Fashion-MNIST, and CIFAR-10 image sets, full replacement collapsed quality fast, matching earlier results, while the fixed-budget approach degraded only partially.
That partial degradation isn't uniform. A mathematical model of the process, based on stochastic recursion, shows some dataset features are fragile and disappear within a handful of generations no matter what, while others are robust and survive over practically unbounded generations under the fixed-budget setup. That's the practical scenario most labs are actually in, since nobody deletes old training data, they just keep sampling fixed batches from a growing archive.
Still, selective isn't the same as safe. Losing the fragile features first just means the damage is quieter, not that it stops.