There's two aspects to this.

The pretraining (common crawl, i.e. the entire internet. Also books and papers, mostly pirated), and the realtime web scraping.

The article appears to be about the latter.

Though the two are kind of similar, since they keep updating the training data with new web pages. The difference is that, with the web search version, it's more likely to plagiarize a single article, rather than the kind of "blending" that happens if the article was just part of trillions of web pages in the training data.

There's this old quote: "If you steal from one artist, they say oh, he is the next so-and-so. If you steal from many, they say, how original!"