One big problem is that Google also lies about dates when you do a 'before' search, so AI spam still shows up even if you filter it out.
Scraping a new dataset without contaminated content is impossible now. There's no authoritative source on the age of any given website. (You can try places like the Internet Archive, but most such archivers aren't interested in handing their data over to AI firms because of the bad PR.)
The only way to ensure you don't have AI content is to use a dataset created before ~2021-2022. This too is hard because all the big ones are illegal to posses as they contain CSAM.
The real answer is that AI developers played themselves. Humanity was creating exponentially more data each year, and by rushing these AI tools out the door, all that data is useless now. By the end of this year there'll be more new "AI-polluted" data on the web made after 2021 than the total amount of "clean" data made up to 2021.