One thing I get asked about often is “model collapse”: will AI development stop as more and more of the web gets filled with AI output, contaminating future training data?
This new paper suggests model collapse won’t happen, as AIs can train on a mix of original human & AI data.