they were; they’re not needed anymore. I trained a few ML models myself and I guarantee they had more than enough data to overfit by the first few months these captchas collected data. A paper from 2019 already showed a success rate of 97% (iirc) of automated models trying to solve these captchas, and trained on much less data. With all the data they collect you can be sure they don’t need any more of it for several years now.