This is only a very rough analogy, but consider these text AIs like lossy JPEG compression for images -- using an efficient statistical representation of the content, you can store a lot of information in a little space and reconstruct something aaaalmost like the original. If I (and a few hundred million other people) have gone through the trouble to create a nice PNG picture or whatever, and then someone comes along and saves it as a JPEG at.. say.. 95% quality, that JPEG will look pretty great. Hard to distinguish from its source.
But what happens when our JPEG saver only has other JPEGs to start with? And soon, is saving copies of copies of... you get the idea. The images lose more and more of their original real-world structure, and become dominated instead by artifacts of the compression and reconstruction processes.
That's basically the (a) challenge for this AI training paradigm of scraping web-scale content. The more that content is generated by the previous generations of AI, the more we're asking the next generation to learn from copies of copies of the original human intelligence that powers them