Model collapse
Decline in a model's quality when it's trained, generation after generation, on AI-made data in place of human-made data.
A model learns the common patterns in its data well and the rare ones less well. If the next model learns only from the first one's output, the rare patterns are thinner still, and over several generations they disappear. Researchers writing in Nature showed this happening: the unusual "tails" of the data went first, and outputs grew narrower and eventually garbled.
The worry is real because the web is filling with AI-made text, and future models train on the web. The research also points to the fix: collapse appeared when generated data replaced human data, not when it was added alongside and filtered carefully.
For example
Train a model on its own stories, then the next model on those, and so on: rare words, unusual plots and minority viewpoints thin out first, until every story sounds the same.
Put it to use
Where you'll meet model collapse in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.