Those “@-@ tailed jackrabbits” in your link made me laugh. Emoticons in species names? Why not?
I think that we could minimise the loss of integrity if the data is “contained” in a way that your typical user wouldn’t see it but bots would still retrieve it for model training.
And we don’t need to restrict ourselves to use LLM-sourced data for that. The model collapse boils down to the amount of garbage piling up over time; if we use plain garbage we can make it even worse, as long as the garbage isn’t detected as such.