The training data in this case is feedback from users - reported responses. It's only logical that as that dataset grows and the developers have worked on it fo

The training data in this case is feedback from users - reported responses. It's only logical that as that dataset grows and the developers have worked on it for longer, the 'toxic' answers will become more rare. And of course, 'caution' in this case refers to avoiding bad PR, nothing else.