Why Model Collapse Isn't a Problem

As AI gets more and more prevalent in today's internet; many have asked themselves, what happens when AI is trained on itself. The way current large data models like chat gpt work is by training themselves using texts on the internet as an example.

But as more and more of the internet is saturated with AI writing, some are worried about AI being trained on itself, causing the model to deteriorate and become nonsensical. A similar example would be making a copy of a copy of a picture. However, studies have shown that not only can the problem be mitigated and nearly eliminated via very basic curating, but even if it is allowed to run wild, the problem would be minimal.