Definition
Plain language
When an AI trained mostly on its own past output gradually loses variety and gets worse.
As stated in the literature
A degenerative phenomenon in which recursively training generative models on their own outputs narrows the learned distribution across generations; a cousin of template collapse and semantic collapse driven by self-referential training rather than policy-gradient dynamics.
Why it matters: It warns that feeding AI its own output can quietly erode quality and variety, a real risk as AI-made content fills the internet.
For example, an image generator trained over and over on its own outputs might slowly forget rare styles and start producing bland, samey pictures.
Heard on the show
“The second implication is about how we think about model collapse more broadly.”Episode 073 — When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving