How are these models, still spouting out lots of seemingly okay but actually non-sensical and contradictory ouptuts going to deal with training the next version

How are these models, still spouting out lots of seemingly okay but actually non-sensical and contradictory ouptuts going to deal with training the next version, once the training set will contain shitloads of non-sense generated by the previous version? As soon as it gets "good enough", it's not going to be easy to tell it's actually bogus. What's to prevent people from poisoning the web with nonsense? Or are these only ever going to be trained with known reliable sources?