Model Collapse
Model Collapse
A phenomenon where flawed training on AI-generated data gradually degrades an AI's output quality and expressiveness.
In Simple Terms
Model Collapse happens when an AI keeps training on unfiltered AI-generated data over and over again -- causing its performance and the variety of its outputs to break down. It's a bit like photocopying a photocopy again and again: each copy gets a little blurrier than the last. The problem shows up when AI-generated text or images are used unfiltered as a substitute for real data, across multiple generations of training.
Behind the Name
The name reflects exactly what happens: the AI's predictive model gradually loses accuracy and diversity, as if it were collapsing in on itself.
Take a Closer Look!
Model Collapse is a phenomenon where an AI keeps taking in data generated by itself or other AIs, unfiltered, as a substitute for real data, and repeatedly trains on it across generations -- causing the quality and diversity of its output to gradually break down.
Data created by humans naturally includes rare, low-frequency features, but when an AI generates data, those rare features tend to get dropped, so diversity slowly disappears.
When that skewed data is fed into the next AI's training over and over, valuable information gets lost with each generation, and eventually the output narrows down to a bland, repetitive sameness.
This kind of collapse happens when generated data is used unfiltered as a substitute for real data and fed back recursively into training.
On the other hand, properly keeping real data intact while mixing in carefully filtered generated data is a technique called synthetic data training, which is actually used to improve AI performance.