I think we should also be aware of the technological reasons for separating AI-produced and human-produced images: New effects/risks can occur when the ML syste

I think we should also be aware of the technological reasons for separating AI-produced and human-produced images: New effects/risks can occur when the ML system is naively trained against its own outputs. It's not hard to imagine multiple companies trying to periodically scrape and build new image datasets and retrain the newest models against the updated dataset. But if stablediffusion vN is mostly trained against stablediffusion v(N-1) images, then it's possible for the sequence to not actually improve even though more data and computation are being thrown at it.