They've found that if you train AIs with the output of AIs then the results degrade. I'm not sure whether they've done this with artbots, nor whether it's about feeding then exclusively on AI output rather than a mixture, but it's possible that they'll have to develop accurate categorisers to avoid this problem.

I can't help thinking this is similar to group-think or internet echo chambers, where groups self-support themselves into counter-evidential mindsets, and then become invested and refuse to examine contrary evidence. It's not exclusively an internet thing, but the algorithm-selected feeds we tend to consume these days do appear to promote that phenomenon.