The sources used for training these AIs are publicly available sources like Common Crawl. If having a copy in RAM is a copyright violation, then there are copyright violations occurring well before any AI ever sees it.