“Researchers warn of ‘model collapse’ as AI trains on AI-generated content.”
It’s funny—I talked about this risk a few months ago, but I didn’t come up with an attention-worthy phrase for it.
The lesson is not about my (limited) prophetic abilities. It’s about the power of memorably naming core concepts as a way to drive your insights.
As for the specifics, yes, there’s a highly probable risk that generative AI will eat itself as it runs out of human-created content to digest and imitate. The research paper itself is quite technical, but its introduction (see below) makes the point accessible.
So, yes, don’t stop writing your own stuff—the AIs need you!
In this paper we make the following contributions:
- We demonstrate the existence of a degenerative process in learning and name it model collapse;
- We demonstrate that model collapse exists in a variety of different model types and datasets;
- We show that, to avoid model collapse, access to genuine human-generated content is essential.