How are these systems trained to make sure they're not polluted with images created by AI already?

I can see how it's semi-easy to have a database of images up to a certain date (before AI generating algo existed), but what about if you want to train the AI with modern data?

For example, say if I want to query, in 2035: "Flying hippos in the style of McGrundsBulle" where McGrundsBulle is an artist who only started producing art in 2025... How is the AI then not going to trained with "flying hippos" generated between now and 2035 by AI like DALL-E 2 / midjourney / etc.?

I'd argue it wouldn't make sense to train better AIs using current AI's generated pictures for many of them are the stuff of nightmares and it doesn't seem to be because they've been trained by nightmarish datasets. They're just creepy due to how the current AI works and many details are obviously off, which makes them very creepy to humans.