I think it all depends on how good our tools to detect AI generated content become. If it is not distinguishable, then the internet is probably about to be flooded by AI generated content which in turn means AI is going to be trained more and more with AI content, degrading the model in the process.